This architecture provides a solid engineering blueprint for scaling AI agents into a coordinated workforce. It successfully replaces the "magic" of prompting with a structured system of orchestration and accountability.
Instala nuestra extensión para buscar dentro de cualquier video al instante
Escaping the UnderclassAñadido:
Hey, we're live on Twitch.
Hey, we're live on Twitch.
I think we're live everywhere.
Are we live everywhere?
Let's check.
We're live on x.com the everything out.
Let's check.
>> We're live on YouTube.
>> Hello. Hello. Hello.
>> Are we live on LinkedIn?
>> X.com.
How would I even know if I'm lived on if I'm Hey, I'm live on LinkedIn.
And then I think I'm definitely live on Twitch.
Hello, Lennav.
>> We're live everywhere, baby.
What's up, party people?
Can you hear the music?
>> Hello.
>> Are we live on LinkedIn?
I don't see >> there's some chat music is on. Thank you, Lennis. Lionus. Lennis.
>> Hello, Lennis.
>> You like the stream already? Oh, that reminds me. We got to reset the goals.
>> What's up, party people?
>> You can hear the music. Good. Okay.
>> Can you hear the music?
>> All right.
Can you reset the goals? Uh the three goals for Yeah, I think I think you're right. I think there is background music playing somewhere.
Uh it's LinkedIn is making noise.
All right, we need to reset the goals for today.
Can you reset the goals on the streaming template today? We're still going to go for 100 likes. We're going to redesign faster.dev and we're going to add team pricing to solo.
Let's see if uh let's see if Codex can figure that out.
Hey techie tunes team mirror. I think I uh muted that tab. Sorry about that.
Hey Joel. Hey Nick.
Mick Labowski.
All right. So I think we're good. I think we're streaming.
Codeex is working on updating. Let's do um faster.dev redesign. We're going to make that what we're currently working on.
Good. Whatever. Thank you. Thank you, Morgan. It is uh 12:47. Good afternoon.
Howdy.
Howdy, Win.
All right. It looks like it updated the Yep. Yep. It updated the goals.
Got my currently building. Got my stream template set up. All right, we're ready to cook.
What's going on where y'all are? Where is everybody? Where are y'all watching from?
We got a bunch of stuff we got to do today. 700 p.m. Not so bad. Nice little evening windown.
I did finish the canban board. Um I don't know if I want to ship it.
That's I don't know.
Hey, Jack from Minnesota. 900 p.m.
I assume that flag means Germany. That's my guess. I'm not a worldly man, but I think I know that that one is Germany.
Did y'all see uh AMP released a new AMP released a new version?
Belgium Vancouver.
What's up, Topper?
Uh, so AMP just released a new version.
Now, this is my logged in.
All right.
Uh yeah, I'm working on it right now. I had a designer do it. Um so it's done. I just need to turn some attention to it.
Ooze 1000 p.m.
Uh where's AMP rebuilt? Read more. This is the new AMP coding agent. One of my favorites. This is the only uh I wouldn't say I have no plans for that.
Hey Sid. Uh this is the only non native coding harness that I use. So I use Claude Codeex. I don't use Gemini so much anymore. It's not very good. Um I think the model is good. I just don't think I think the model is smart. I just don't think it's good at calling tool doing tool calls. Oh, you'll see a redesign. Well, that's what we're actually we're going to be working on it. Oh, you mean a redesign of solo.
We're getting close, but this is the new We got to figure out what's new about AMP cuz it just came out today.
Uh, I got two concurrent viewers on LinkedIn. I'll be honest, if you're on LinkedIn, I don't know if your comments are coming through because I've never streamed on LinkedIn before. Um, MCP is so powerful. Sub agents with to-dos. Yes. Yes. Yes. Hell yeah, Andre.
Yeah. Sorry, Google.
I saw Claude just announced they're doubling uh I think they're doubling rate limits. I don't think they're doubling like I don't think they're doubling usage limits. I think hey thank thanks Sid.
I do have all the platforms.
Yeah, I don't know that uh I don't know that a lot of the middle managers on LinkedIn are watching. And if you're on LinkedIn and I just insulted you, sorry about that. Um, so I don't know that I don't know, Claude.
I think they just doubled us uh rate limits, not well that says usage limits. Okay, so today we're doubling the 5-hour rate limits.
We're removing the limit reduction in peak hours and we're raising API rate limits.
Uh, I've not tried Polycope.
So, I don't think that like your max plan suddenly gets 2x usage.
I think you're just going to hit the 5 hour cap less often.
Polycope is not an official Laravel tool, by the way.
John, it works. I see that.
It works.
So, yeah, big things happening. Um, yeah, sorry, Sid. It's the middle of the day here. Hey, Taff. It does seem like weekly limits are the same. That's what I'm that's what I'm gathering, but I think the 5hour limits are maybe more generous now. Uh, has anyone here used AMP?
Do y'all use it?
AMP is one of my favorites and they just rebuilt it. Uh the AMP guy, one of the AMP guys, Token Town is still on, God willing. Uh one of the AMP guys is coming to Hey, who's that?
Um this guy right here, he's coming to Laracon. He's going to be a speaker.
AMP is nice, but it's spendy, man. It's so expensive.
So, if y'all aren't coming to Lar Laracon, you got to come to Laracon.
Uh, yeah, AMP is awesome. So, the two reasons two reasons I love AMP. Um, and this is one of the things that's coming in the AMP faster.dev course, so uh stay tuned.
Um, two reasons I love AMP. One is the Oracle. So, it has these built-in tools.
One is the Oracle and one is the librarian.
And the Oracle is like basically deep research for your own codebase. And then librarian is like deep research uh for open source code. And so it can go out to GitHub. I don't know how they do it.
I mean, obviously every model or every harness can search GitHub, but AMP is doing something crazy. Um, because the way that they search GitHub is incredibly efficient and incredibly detailed. Um, AMP comes from source graph. They were spun out of source graph, which is like it makes um it like helps you get a handle on your actual your own codebase.
So they were working on this kind of stuff for a super long time uh before agents came along and then AMP was spun out of source graph. So I have to imagine they're using some source graph magic under the hood at AMP. Uh but that's why I like AMP. I bet they probably do still have access to the source graph index. I bet you're absolutely right. So what's new in AMP?
Remote control. Everybody's doing remote control. Man, this is cool.
Remote control is coming to solo, by the way. So, one thing that AMP has always done is they've made you do handoffs.
So, they never had um auto compaction and you've had to hand off like when your context ran out and they have decided not to do that anymore. So, they're going to do auto compaction now, which is cool. Uh you don't on AMP, you don't select a model. Um, you select a mode.
So, this is the new amp. And you see over here it says deep two. Well, now it says large.
Oh, look at that. You can click on it and change it. That's new. Large, rush, smart, deep, too.
Hey, we got to figure out what large is.
Smart and rush and deep have all been there. I don't know what large is.
Hey, Jamie.
Where are you these days, Jamie? I forget.
It's not 11 Labs, is it?
Remote control is so overrated. Yeah, I'm not I'm kind of torn on it.
So, they have a new plug-in API.
Hey, Adrian. Yes, I read your DM. No, I have not fixed it yet. Sorry. Um, they have a new plugin API, which is nice.
Dang.
So, you can register your own tools.
That's kind of cool.
Lots of people are going towards this steering thing.
Yeah, you choose the type of assistance you want and not the model and they figure it out.
So queuing versus steering is interesting.
Queuing is like, "Let's do it after you're done." And steering is, "Hey, you're you're going off on on the wrong direction."
Permissions are loosened. That's nice.
Damn, look at this.
79% less CPU for a 2e.
70% less memory that rules. Good job. Handoff is gone.
That's what I was talking about earlier.
So now instead of um Windows support implementation, not yet.
We're getting close. I think the Windows PR is actually passing. So it's just a matter of me being brave enough to do it. Um handoff is gone. Compaction made it obsolete.
Okay, so they have auto compaction now just like the other ones.
They removed all the themes and they removed manual bash invocation.
Interesting.
53 codeex fan.
Nice.
Uh let's check that. Let's check that.
Uh PR Look at that.
Look at that. Oh, cool. The zoom works.
Fix tar dev on Windows. Passing.
So, we're close.
How close we are, I'm not entirely sure.
Um, the campaign board is there. I don't like it. Um, I don't know if I like it.
Uh, any update on option key meta key issues? Not yet, but I promise you I'll get there. It'll be in the 66 or maybe I'll do a 70 release. Um, all right. So, we didn't find out what large is.
Wonder what large mode is.
They have a pretty good uh owners.
Deep large mode. What does large mode do?
Oh, this isn't new. I just didn't know about it.
Um my tar experience has been good. I will show you I'll show you the rest uh JSTS breakdown large mode. Okay. So it's just the million context window.
Yeah, I don't know that I would use large mode now that they have auto compaction. I'm just not sure I would use it. The large mode.
Um, all right. Solo. There you go.
50% 55% uh 55% TypeScript 40% REST.
And a lot of that TypeScript is um tests and harness harness for the tests.
And look at this.
Isn't that freaking wild for one person?
That is insane to me.
I think initial commit was five months ago.
The Swift is for like native pop-ups and stuff and a little bit of like um native desktop notifications, that sort of thing.
The performance focused parts, I wouldn't break it down by performance and non-performance. I would break it down by front end and back end. Um the front end is all React and the back end is all Rust.
And the reason the reason I went with all my interns with all my freaking US dollars um the reason I went with Tari instead of Electron is like this right here. I mean this is streaming this is streaming bytes from the back end to the front end and there are a lot better libraries on the REST side. Uh a lot better libraries on the Rust side for handling processes pty uh vt 100 screens that sort of stuff. Um, I have never written a line of rest in my life, including up until now.
So, there you go. Look at that. Solo, good job, Solo rendering the TUI with dragable. Look at that. You can drag it.
You can click it. You even get hover states. Pretty proud of that.
Oh, Morgan.
Yeah, I chose React because of AI, which is so sad to me because I think Vue is better. Um but like um Pierre Computer Company but like these people um where come on guys diffs.com got to have a link there my friends but like this this is React And if I want, which I do, to have a diff viewer, I'm going to use the best one in the world and it's this one.
And it's React. And that's just one.
That's just like one example. Everything is React, which is to me frustrating because I prefer Vue. But since I don't write the code anymore, um it's not up to me. All right, we got stuff we got to do.
Um, how for the GitHub hub diffs? What do you mean how for how long for maybe is the missing word? Um, I don't know. I need the I need the redesign to land first.
Local first, baby. SQL light.
Um, okay. So, uh, any tests on Linux yet?
I have GitHub workflows that run on Linux. Um, but I have not actually launched it on on Linux yet.
All right. So, we're going to do a couple of things. First, you got to like the stream on YouTube because last time we barely hit 100 and it was almost devastating. Um, so he never took the time to finish the sequel course. Oh no.
Well, that's pretty common these days.
Okay, so we're going to be jumping around. Okay.
Um, here this is desktop one, two, three, and four are all solo.
Um, guys, you got to finish the courses.
Actually, I don't believe that you don't have to finish the courses. You need to watch as much of the course as you want.
That gets you uh back to doing your job, hopefully better, and hopefully making more money.
Um, are we going to get a fixed schedule for the lives? I love being surprised.
Um, give me just give me a second to like get into the habit of going live and then I will commit to a fixed schedule. I'm just, you know, I'm just trying. I'm trying my best here. We'll get there. Stick with me. So on desktop we are working towards um what's the scratch pad we've been working off of on desktop. Hey thank you Josh.
Mr. Grumpy Monk I'm live. Are you vibe coding my friend? All right so on desktop uh we've been on solo we've been working towards this scratch pad right here. canonical command surface generation.
So, Solo has a pretty robust MCP. What it doesn't have is a robust CLI or a robust uh HTTP surface area.
And so what we're working towards is uh making all of those paths share the same underlying commands so that Solo can become uh the most orchestrable tool that you use. So you could do it via CLI, HTTP, MCP or eventually what is this a JavaScript sandbox? No, say no more. Um, okay. So, that's what we're working on over here. CLI could be better.
Yeah, YouTube I I pump more I pump more bits to YouTube. So, YouTube is always going to be sharper and it's likely going to be faster or like less less laggy. Um, okay. What remains on that scratch pad?
All right. All right. Then we need to fire up a faster.dev agent and we're going to start npmdev.
Faster needs a little love. So I've had a designer working on faster.dev and we're going to start to incorporate it.
Oh, hey Josh. Co-working together.
Same AirPod Mac. Yeah, I just got these AirPod Maxes.
How exactly do you make the solo work without making you do anything?
Uh, I'll show you in a second.
Uh, Will might be right. Word on the street you're shipping more features than your designer can keep up with.
There's a chance Will is right because he is the designer, so I'm inclined to trust him on that front. Um, all right.
You want to see a little uh a little orchestration loop? Um, okay. Let's see.
Where could I go?
You could, somebody some people have created skills for solo orchestration.
Um, and that's useful, but you don't necessarily need them because you can use uh you can use um solo timers to make sure that it stays awake. So, you could do something like this. Let's say, let's rename this one the orchestrator. So, this is going to be this is going to be your lead agent. Uh, does Solo use agent harnesses? The good thing about Solo is it's a meta harness. So, it sits above your favorite harnesses. So, you can use Open Code, you can use um you can use Pi, you can use Claude, you can use Codeex, whatever.
Yes, solo is a harness for harnesses that harness the harness to harness harnesses and harnesses. That's that's better than I could have said it. So, what you can do is you can do something like this. You can set up an orchestrator and then you could get clawed out and you could turn this one into a watchdog, right?
So, what I would maybe do if I wanted it to run overnight, I would do this. So, I would copy the link to the orchestrator and I would say, "Can you see that process via the solo MCP?" I know that it can, but we're just going to ask.
We're meta harness mogging. You bet we are.
Extra high, baby.
GPT 55 extra high all the way.
I don't know, Josh. I was pretty proud of that. I was pretty proud that I got that. Um, how are they connected via the solo MCP?
All right, so I just spawned uh a Claude and a Codeex. This is uh somebody asked how would I do like an overnight loop with Solo and this is just an example.
Um, so I spawned this Claude and I spawned this codeex. Now I named this codeex orchestrator. And what I would probably do is here I would point this uh I would point this codeex to a scratch pad. I don't think we have any scratch pads in faster. Yeah, there are no scratch pads in faster. Um, so I'll just give you an example. LLM consciousness consensus. That's an interesting one. So I'll copy that link and I'll say something like, all right, you're the orchestrator. Your job is to never write code. What you're going to do is you're going to work off of this scratchpad and dispatch tasks to other agents. Prefer to work in parallel when possible. Use the solo MCP to spawn them. give them their tasks and when they're done uh you can harvest them, check it for QA, uh and then close them.
Boom. And then I would paste in the scratch pad right there.
One harness to rule them all. That's right.
Den I don't know. Chat GTA 345. It's a cool name. So this is the orchestrator.
So this is the one. Um in fact, we have an orchestrator running down here. So, we can do this for real.
All right. Go ahead and work on 959 then.
Now, I told it to go ahead and work. I don't know if it's going to spawn a sub agent. Um, but this is working off a set of to-dos.
And so, it's just working through a whole plan that we've worked on together.
Um, so this one is looking at to-do 959.
So up here I've given it this uh I've given it this task. Now the watchdog, what I would do is I would say something like, "All right, I want you to use solo MCP timers to um just keep an eye on the orchestrator. If it ever runs into permissions prompts, I give you uh permission to approve anything that uh it needs to work towards its goal. And if it stalls out, I want you to send it some input to tell it to keep going. You can set a 30 minute timer that wakes up when the orchestrator goes idle. Um and if it's still working, you can just let it keep working.
And then I would send everybody off to the races. That's all I would do. The solo timers are interesting. Um because watch, we'll do uh we'll do a new Claude.
Can you see the solo MCP?
I'm glad y'all like the tunes.
Solo MCP is available. What can you do with the solo MCP?
I'm going to show you the timer setup because I think it's really interesting.
So you can, this is, this is the real powerful part. It can spawn processes and talk to them.
Um, I want you I want to test the timer.
So, can you set a 15-second timer for yourself?
Hey, Naples Nola, welcome. Glad you like the tunes, man.
All right. So, this one just set a 15-second timer for itself. Look down here.
See, it's counting down.
Five, four, and this claw is totally stopped. And my hands are off the wheel.
And now Solo. Boom. Solo's going to fire the timer.
There you go. 15 second test worked. Um I have I considered creating a dedicated uh a dedicated orchestrator agent? No, because I'm super focused on building the right primitives so that people people can build their own agents because what I find lacking in a lot of these other tools is they want you to work they want they force you to work in their style and then you know what everybody's got a different style. So, I'm focused on building the correct primitives so that people can work in their own style.
All right, we got to get to work. What are we yapping for?
Um, can you go ahead and merge this branch just locally into main?
Yeah, exactly. Um, what about my question? Switching based on rate limits. So, right now you could do that. That's complicated because you'd have to figure out what the rate limits. you'd have to have an agent spawn a new agent and introspect what its rate limits are, which would be doable. Um, but a little a little bad. So, my thought is, and I'll be curious your take on this. My thought is, you see how I have codeex bar up here? This is just an open- source tool.
Codex bar exposes all of these rate limits nice and cleanly. So, you can see I have two codeex plans. One of them I'm already in deficit uh and one of them I'm still doing fine. Codex bar uh you know it works with Claude and Gemini and all that. Codex Bar also exposes uh a small HTTP server and so I'm thinking I could allow you to say like or I could see if you have codeex bar or whatever the other one is installed. I forget what the other menu bar is, but I could allow you to say like, hey, I've got this installed.
Therefore, in the Solo MCP, expose tools that tell me how much usage I have. And then you could tell your orchestrator agent, hey, just call the solo MCP, make sure we have enough usage. If we don't, fall back to this other one. So, yeah, we could totally do that. Uh, we're just not there yet.
All right. So, you are still cooking.
The other thing we need to do is for solo term. Um, all right. I'm just going to So, we need to We're doing three things at once cuz that's how we live these days.
Um, scratch pads were inspired by my frustration with random random markdown docs in uh in my repo because I always like I'll do deep research on the web and then I'll have to drop it somewhere and then I have to remember where it is and now they're just all they're just all scratch pads.
I just need to make sure that nobody's like chatting on LinkedIn and I'm ignoring them.
Yeah, there there are two viewers on LinkedIn. So, I think I think we're going to be okay.
Yeah, I think it I think having uh limit via MCP would be very cool. Um, okay.
So, we need to do some brainstorming here.
Uh, we need to do some brainstorming here. I need to offer I need to offer team plans for Solo. Um, and right now we're just not set up to do that. So, I've had a few people contact me about enterprise uh enterprise licensing and I think primarily what they're looking for is ease of billing and maybe like seat management. And so right now every license um is just basically tied one to one to a Stripe checkout.
So at the very minimum we need to add new uh new Stripe products. But then we also need to add a seat kind of like a seat management interface. Um oh goodness I actually don't know if we need seat management.
Um, that is a good question. So, I need you to help me think this through. I need you to challenge me on some of these things because um, Solo right now, you don't have to log in at all to use it. The only time you ever have to log in is to get a license key.
And so if the manager, let's say the purchasing manager buys 50 license keys, do I need people to log in? I don't know. Um, do I give the manager 50 license keys?
That sounds kind of annoying. Do I give them 50 oneoff coupons? That sounds even more annoying. Um, do I gate it based on domain?
I don't know. That's a good question.
So, why don't you start asking me some questions and we can narrow this down.
All right. Go solo with a team. I thought um solo together, but I don't know. I I may use solo together for something else.
All right. Faster.dev merge.
Okay.
So, you're merged. Now, we need to find we need to find the new repo.
Where was that repo that I was invited to?
Sidecar Laravel 13 support bane of my existence is open source. Um, where's that repo that I just got invited to? Probably in my email.
Solo plural. That's pretty good.
Uh yeah, extra high fast, baby. Sometimes, you know, when the market's moving, sometimes you just got to pay the price.
All right, here is the redesign repo for faster.dev.
Um, okay. I have a new design for faster.dev. And here's the full repo for it.
Um, I want you to clone this into a temp area so that we can uh start uh bringing over the design. But before you do anything, clone it down and let's take a look at it. Uh cuz we need to come up with a good plan before we just like start doing onesie 2Z.
1 Z 2Z. That's not really what I meant.
Single use license. Yeah. I don't know, John. That's a I I don't know. Um I don't know. That feels maybe feels a little bit annoying to me. Um but I could, you know, I could be wrong.
Maybe that's what Maybe that's what they want.
All right, I'd start by separating team plan into two different jobs. Buying, one company wants to pay once for in people and enforcement, sign, revoke, and audit. Yeah.
When people ask for enterprise licensing, what have they specifically mentioned?
1 Z 2 Z. Um, can I buy 20 license? It's definitely the ladder. It's not SSO, security review, anything like that.
The company buys 50 seats and 58 people use it. Is that a serious problem? We been have serious problem. Uh, honestly, it's not a serious problem.
A seat is a named human email address device concurrent. It is an email address, I think.
Does the app currently require an email tied to the license key?
I think you can just use a license key, but they are all tied to an email.
Licenses cannot be revoked or rotated.
I don't think buyers need to know who is using solo right now.
This John John Sugar, this is what I was thinking.
I mean, God willing, we get 500 seat deals. Um, that would that would be awesome, but that would be super annoying. I think they're mostly small.
I would not start with coupons or 50 one-off keys. Yep, I agree.
I don't need revocation.
I think it's mostly centralized billing.
I don't need SSO probably.
Oh man, I probably don't need revocation and assigned seats to start.
I like the idea of the um I like the idea of the billing manager um the billing manager buying a certain number of seats and then having a like uh like a link where they can just drop it in Slack and anybody that uses that link can go and claim one of those seats. I think that's the shape I like the most.
Well, Jerome, that's a good idea. I wonder how how do I hook up live now notifications to Discord?
So, you're still cooking. Oh, you're trying to launch Chromium.
All right. What have you done?
If you have not already liked the stream, please go like the stream.
Please help me get to 100 before midnight.
Okay.
To-do 959 is complete.
Is there anything left on that scratch pad or to-dos that have come from that scratch pad that we haven't yet completed?
Does anyone know how to add um live now notifications to Discord?
All right. So, the faster redesign is cloned.
Stream cord is the easiest. What about treating it like testflight in invites billing manager?
Thank you, Adrian.
Yeah, there's a Discord solotterm.comiscord.
Uh, I need to add it to faster.devisord as well.
I appreciate y'all y'all Twitch viewers so kind. Um, okay.
Yes, I agree. So, this new design, this new faster.dev design is a radical departure. Um, it's also built in a completely different stack. So, I had a designer design it and then they built it out.
Um, and now I'm just going to yoink. I'm just going to yink it over. Um, so let's see if we can get some multi- aent orchestration going. All right. I want you to um create a scratch pad with the entire plan. Um, I agree. I don't want to I don't want to pull I don't want to pull in their tech choices. Um, I just want to pull in the aesthetic design while keeping our Laravel inertia setup. Um, and I expect this to be a pretty complicated surgery to get all this ported over. Um, so I want you to write a full scratch pad with um the ideal uh like timeline or flow that we should do this in.
um with checkpoints along the way.
You switched from YouTube to Twitch.
Okay. So, tell me, why do you guys like Twitch? I've never gotten I've never understood Twitch.
Cool. Seems like we're done.
Heck yeah. That's awesome.
Okay, why don't you spawn an amp to review this for now?
Better chat, better with fun stuff. It's where the cool kids hang out. Yeah.
Yeah. I feel like YouTube So, I feel like Twitch is for like the points and the hype trains and like the emotes and all of that stuff.
Um, and I just feel like I'm too old for that. Is that crazy?
Don't have to use your real name. Oh man, I love using my real name on the internet. Using my real name on the internet has been one of the greatest uh one of the greatest hacks for getting ahead cuz everything I've ever done is tied to my real name.
All right, so this one is sending. Boom.
Look at that. Now we're using AMP.
We're using AMP to review all that work that Codeex just did.
Got a little sub agent going. You see, look, here's the notification. It set a timer for itself. I set a timer to wake me when AMP goes idle so I can harvest without polling. Hell yeah, brother.
Prime is uh Prime is too old, but he's too young at heart. I'm both old and old at heart. So, Prime Prime still has that youthful energy.
Me e6bot. It does YouTube go live notification. Cool.
The Claude code limits are 2x. I think I think that's not quite true. Um, I think the 5hour rate limit is 2x, but I think your monthly or weekly plan is still the same. I think that's what we decided.
All right, that was um, no, use the solo MCP. Don't write it to a markdown. See, this is where skills would actually uh this is where skills would be helpful. But I'm purposefully not using skills myself because I want I want the I'm trying to make the MCP as good as possible without any help. Um, see, just write this random markdown document. That's the thing I don't like.
Okay. No existing scratch pad. That's right. You're creating a new one. That's correct.
Um AMP is uh AMP is its own harness. And so with AMP, you can't like you can't choose the model. Uh it uses whatever model it deems appropriate. But I find it to be incredibly good at using uh tools to do things. So they have a couple of tools, the Oracle and the librarian. And I feel like they're just really really good at introspection and actual for the librarian open source research.
All right. Solo, what do you say?
Do not make the Slack link itself the license. Make claim link that mints individual licenses. That's nice.
Billing manager buys solo team 20 seats.
They get one reusable claim link. Anyone with the link can claim a seat.
Claiming creates an individual license for that person. The app still only needs a license, not a login. I agree.
Billing stays centralized. I agree.
Yeah, I'm with you, Derek. All of these things start from a blank slate and they never remember anything.
V1 could be very small. Organization.
Yep. Yep. Yep. Team subscription. Yep.
Yep. Yep. Claim link. Yep. Yep. Yep.
All right. So, we're still trying to figure out solo team licensing here.
Should the claim link be open to anyone with the URL?
I think it should be open and optionally restricted.
Um, aren't MCP servers more token heavy than skills?
Um, no they're not. So, skills do just load the description up front. Um, MCP. So, depending on the agent or the harness that you're using, they can have like deferred tool you or deferred uh tool searching or whatever.
I think Claude does deferred tool searching. Um, but it just loads the descriptions up front as well. And for Solo, all the descriptions are incredibly small. And then there's another there's a help tool that's exposed that the agents can call to get more detailed help. So a lot of MCP servers will shove everything into the tool description. I don't do that because it does get bloated and so the tool descriptions stay super small.
All right. I agree I think with all of this.
Um I think the claim link should it should meant an individual license. I agree.
Um anyone with the link can claim a seat.
Claiming creates an individual license.
The app still needs just a license, not a login.
and billing stays centralized. I agree with all of that.
I don't even know if we need a new data model for organization because an individual like the billing manager could just go to the homepage and buy 20 seats. They don't have to create an organization or anything like that. Um, yes, we'll do a Stripe subscription with quantity. That's how we'll that's how we'll do it on the Stripe side. Um, it should, you know, for now we're not going to have domain gating. It's just going to be It's just going to be the link and they can just go claim it.
Should a claim link be open to anyone with a URL? Yeah, for now. We could potentially do domain gating later. Once someone claims a seat, do they Um, I think they can just get I think they need to log in and then they get the license key displayed in their dashboard just as it currently is. If someone bought 20 seats and the 21st person clicks, um, yes, show all seats claimed.
The billing manager can increase or decrease selfserve.
I think Stripe customer portal is fine and yes claimed team licenses look exactly like normal licenses. Yes. So what is that with all of those answers in mind? What does that data model look like to you?
All right. How's the burn rate on fast mode? It's insane. Um that's why I turned it off. It's like double. I love how we're using AI but know actually how your app works and doesn't work and remember the choices made. Boy, if I didn't, I would be in a world of hurt, my friend. Yeah, I think I think um there's a lot of conversation about everything, but I think there's a lot of conversation about like handwriting code, and I have no problem not handwriting code. What I have a problem with is when you have no idea how the thing works and you're just layering on more if statements to make the new thing work. That's I think that's where people get in trouble is they're like, I don't know, make it work. And then the model wants to please you and it's like, all right, well, I can add another if statement here and hopefully it'll work.
and it probably will for a while, but then you get to the point where you're like, "Dude, I've got 50 if statements because I didn't come up with a like a theory of mind of how this should work."
Hey, Joe.
Um, all right. AMP reviewed and found no mustfix issues. No must fix, no medium, later cleanup.
Keep this is as is continue later kettle expansion. Great.
Okay. Have we already run pre- push on all these changes?
All right. So, what is what do you think about? Look at this. Look at this.
This screams orchestration to me. Hell yeah.
Look at all that good stuff.
All right, you guys want to see an orchestration loop? Let's try it. Let's rename you rename you orchestrator.
And in fact, let's start with some to-dos.
Um, okay, that's perfect. Let's keep that uh let's keep that scratch pad as the working document, but can you break it down into discrete to-dos and then add the blockers because we're going to start dispatching multiple sub aents to pick them off. Um and I want to have a good graph of what relies on what.
Hey, thanks Jake.
Yeah, this is how I work. I work in solo and I use monologue as my uh speech to text and just hop around getting stuff done hopefully.
Agentic engineering. Yep. Yep. Yep.
Agentic design engineer. That's a good term.
All right. So, Codeex is currently putting together some tags over there.
Let's see what Solo's up to.
X high X high all day. I would still add a teamish model, but I would not call it organizations yet. You need a billing container for Stripe purchase seat quantity claim claim claim license pools team subscriptions or seat pools. my preferred model. Guys, this is the important part. If you get the data model wrong, you're hosed.
Um, you can just do this with Claude.
You can do this with Claude or with um Kimmy K2.5 or whatever.
All right, users. All right, so we already have users as ID email license pools. That's interesting.
That is very descriptive. I mean, it's totally accurate. License pools, an owner user ID. So, that links it back to the user. Stripe, stripe, stripe, stripe, stripe status, claim, claim, and then licenses. So, now licenses has a new uh license pool ID which is nullable.
source I think becomes I don't think source is I don't think that's right because it's you already know whether it came from a license pool or not.
It's true Derek. You get the data model wrong, you're in for a world of hurt.
Okay, that seems right to me. I will challenge one thing on the license pools. I'm sorry, on the licenses table. I don't think we need a source because I think that is duplicative of the license pool ID.
So we're going to add a nullable column.
All existing licenses are by their very nature individual and any future ones that come from a license pool ID are by their nature a team. Um, so I don't know that source gives us anything.
No, man. I'm doing my own live stream.
I'm competing.
All right, Joshua. That's what I like to hear.
I got it right for the time and then it was wrong.
Mac, I'm not sure what that's in reference to. Um, source is redundant.
Hey, we're still smarter than the machines, at least for today. I'd skip it for V1. So, license pools? Yep.
Licenses? Yep.
Yeah, I don't think we need claim that.
Yes.
Yep. I agree with that. Uh final data model. No source and no claimed that.
Why don't you spawn an AMP and describe what we're trying to do and run this data model by it and see if it sees any uh edge cases or holes that we haven't thought of.
I don't know if I've told this one about solo MCP.
No.
Use the solo MCP to spawn an AMP agent to ask it about this data model.
See, now that all these harnesses have their own sub agents, um they always want to reach for that.
Well, claim that um the license only exists if it is claimed because when they click the link, the license will be created and so literally created at is claimed at in that case. So, yeah, John, you're right. There's definitely room for uh there's definitely room for more team features as far as like agentic coding goes.
Here we go. It's off to the races.
Yeah, revocation is a good question. I don't know, but it doesn't have anything to do with claiming that I do know. So, we might need a revoked that, but I don't think we need a claimed that because if it's a team license that's been created, it is by nature claimed.
All right, I've been drinking too much Celsius. So, y'all watch that cook while I use the restroom.
Hey, hey, hey.
sharing choices maybe scratch pads. What is it?
And sharing choices could be nice and maybe choices use something like scratch pads to solve for I'll leave you you two to discuss amongst yourselves. Um, okay.
Let's read.
Oh, that's a nice idea. AMP has the idea to rotate a claim link. That's nice.
So, you could you could change it uh if you know one of your dumb employees leaks it or something.
Unique constraints yada yada yada.
Yeah. Here's if you never revoke in v1, count all pool licenses. Decide now because future revocation depends on it.
Yeah, we got to solve Revocation or at least have a plan for it.
Uh Joshua Berios, you want to see a good use of a scratch pad? You want to see a good use? I'll show you a good use. Look at this. Here's a scratch pad. Faster dev redesigned. Uh, so I had Codeex do a bunch of research on how we're gonna I had a designer create a new design, how we're going to basically rip that apart and bring it in. So, it cloned it into a temp directory and then it examined it. And this is the scratch pad for so you can see all the phases over here. Um, this is the scratch pad. So now, um, it doesn't have to remember all that. It wrote it all down, right? So, we're good. And so I can spawn new agents with fresh context that don't know anything about anything and my north star is still uh preserved here.
So then what I did in faster dev is I told it it's going to be a long way back. Um I told it create a bunch of to-dos based on that scratch pad. Um and so it created let's go to faster to-dos.
Here are all of the to-dos needed to implement this new design. And you can see, for example, this one is blocked by three, but it's blocking one. So, all of these to-dos came out of that scratchpad. So, we did the research, we wrote it down, then we had the agent turn it into to-dos and set up kind of the blocking graph. It's basically memory. Yeah.
Um, so now in faster, I'm going to say, "All right, we're ready. Here we go.
Okay, you're going to operate as the orchestrator. You're not going to write any code yourself, but what I want you to do is use the solo MCP to spawn as many codeex codeex agents as possible.
um you need to make sure that they can work in parallel and not stomp over each other, but spawn as many as possible to start picking off the to-dos and working towards the uh north star as defined in that scratch pad. um every I don't know every let's say after every um yeah after every commit you can spawn a you can spawn a claude code to do a QA review and make sure that we're on the right track. Uh and you can use solo timers to fire when your sub agents go idle. You can use fire idle when any or when all. Um that way you can keep an eye on them. If any of your sub aents need permissions uh prompts approved, uh I authorize you to just send the bites to approve those permissions. Can you say all of that back to me so I know that we're on the same page as far as the plan goes?
All right. So, we're about to see a bunch of agents spin up. Yeah, shared scratch pads for a team. I know. Just had my orchestrator use a scratch pad to write an RFC. Spawn another agent to check it to do sends to another agent.
Yeah, exactly.
Um, so help me understand. Um, I think so. This is the same question as should we pull in to-dos from linear?
This is that same question. And my thought is uh my thought is that like all the to-dos and scratch pads in here are for implementation, right? So maybe I pull a linear ticket and then I'm like, "All right, how are we going to do this? Let me create a scratch pad. Let me create all the to-dos to do it."
Boom. All behind this one linear ticket.
And that's kind of how I think of scratch pads and to-dos as like implementation details, not uh not like long running, you know, repo knowledge, if that makes sense.
All right, let's see. I'll operate as orchestrator only. Correct. I will not write application code. Correct. I'll use solo as the working system.
Scratchpad as the north star. Solo's to-do. Yep. Spawn codeex. Yep. spawn as many agents as is practical. This is why we don't need uh this is why we don't need work trees because it's the orchestrator is going to handle it. It's going to say we're working on one thing first of all and the orchestrator is going to handle making sure everybody is in the right spot. Require agents to lock their to-do respect file ownership and report change files. Yes. Spawn a QA correct. Solo timers correct. I'll still respect the platform's hard approval boundary. Yeah, I know.
Current starting point. Go.
Yep, that sounds good. Get started.
All right. And we're off. Should have turned off fast mode for that one.
God, this is so freaking awesome.
Yeah, SOPs. Yeah, I feel like SOPs may need to live in the repo.
So, look, it's creating comments on the to-dos and it marked one as done. Here we go.
Redesign 01 audit mapping.
We're off to the races.
Man, terminals are hard.
Yeah, John. I think they will have to log in to get the license. I'm pretty sure Joe, you are my number one fan except when you're giving me a hard time, but you're still my number one fan. There's this guy Chirkham on YouTube um that I've had to hide from my channel and I've had to hide from the Mostly Technical channel because he takes every opportunity possible just to be mean to me. And so I just had to hide them from everywhere. Like you don't have to like me, but just go away.
Um, all right. So that is still working.
All right. So codeex read AMP and let's see. AMP agreed. The model is reasonable with a few hard requirements. No source.
No claim that. Yeah. Hell yeah. I did that. That was all me. Well, I did source. It did claim that.
make claiming transactional. Yeah, that's fine. I mean, I don't know. He just doesn't he doesn't like me, I think. Um I think honestly what I think it is is I think he doesn't like AI coding and because I'm AI coding, he wants to give me a hard time. And I get it. Like it's a scary time. So I don't I don't fault him for like pushing back on that. What I fault him for is being rude. I don't have any patience for that.
Uh constraints and indexes, yada yada yada.
Yeah, we'll deal with revocation when it comes.
Yeah, if you lose somebody's kid, you probably can be disliked. But we found the kid, so that's good.
What a story, huh?
What a freaking nightmare.
Classic drop and goer. Uh, what's the best way to ensure orchestrator agents don't approve sketchy commands?
It's a good question. I don't know. Just hope that your agent is smart.
I'm a little off center, but I don't want to move. Maybe I'll just fix it in Maybe I'll just fix it over here because I don't I'm a little off center.
This is uh this is my OBS. Pretty cool, huh?
So, how can I do this? Can I just move you? I'll just move you a little bit that way. Expand you. Hey, there we go.
There we go.
This is running on a tale of two grandmas.
Oh yeah.
Can you drag OBS in a circle? Hell yeah.
Let's have a little fun. Huh?
Um, this is my Mac Studio. Look, this is a completely discreet computer. How cool is that?
So, something that is cool that I'm doing. So, you'll see I'm streaming, but I don't have um I don't have the OBS multistream thing plugged in. Like, I don't have that uh I don't have that plugin, but instead I am running Oh, your poor wife. Um, look. So, instead of multistream, I have uh there's my there's my LinkedIn stream.
So, you can stream to LinkedIn on my behalf, I guess. Um I have this little server that's running on the Mac server or on the Mac Studio.
I have this little server that's just a uh it's just an EngineX server, but it has the EngineX RTMP plugin installed. And so from OBS, I stream to localhost, whatever it is, 1935. I stream to localhost 1935. And then engine X uses ffmpeg to re-encode it and multistream it out to all the platforms. So I got rid of reream. I don't have to use that janky uh I don't have to use that janky OBS plugin.
And that is because of uh what is it? Engine X in Gen X RTMP module.
So this is uh I don't know if it's this exact one.
Oh, yeah. I guess it is.
So, look, here's what you can do. How cool is this? So, you can just set up a uh you can just set up a little EngineX server with an RTMP block listening on 1935, which is what I just showed you.
And then uh this is exactly this is exactly what I do. I have uh transcoding so it's live and then this is the transcoder. And so for each different uh for each different platform well for each different bit rate some platforms use the same bit rate for each different platform it gets ffmpeged and sent out to that platform. That's why YouTube is much higher quality than Twitter or Twitch because I can YouTube supports it and I can send it much much higher quality and I don't have to pay reream and I don't have to use that janky uh OBS plugin. It's super cool.
Yeah, this is that same this is that same uh thing that I live streamed a long time ago.
That's back when I was like, I'm gonna be a reream competitor. There's no part of me that wants to be a reream competitor. I don't want to I don't I don't know. It's hard. It's hard to build things. It's hard to productize things as we're seeing with Solo. It's easy to build things. It's hard to productize things.
Uh my take on Warp going open source. I would love to have $75 million. That's my take on Warp going open source. I don't know what they're doing. Um, I think Warp is very good. I don't I don't know what their plan is.
All right. All right. Why don't you capture all of this into a uh scratch pad with solo MCP and then um turn it into a set of linked to-dos and we'll start working on them.
All right. So, you are done, which means your timer file fire fired and you woke up. And look, it's adding its own work trees. That's fine. If the agents want a work tree, go ye with God. I just don't want to do it myself.
Yeah, I would take 7.5 million. I mean, I'd be barely scraping by, but I would take it. Look, orchestrator. Boom, boom, boom, boom, boom.
Look at all these.
So, the orchestrator has decided now that we're past the first to-do, we can spawn four others and work in parallel.
That's nice. I don't think warping is the plan either. That may very well be true.
So, these are still waiting on their instructions from their overlord.
This one's about to get its instructions. It's always fun to watch this stuff automatically happen. Boom.
Done. Um, and I'm realizing more and more what this solo architecture enables because I sit between, you know, because I own the pty handle.
It's as simple as one agent says, "Hey, I want to inject this into another agent." And the rest back end says, "Yeah, obviously I'm already proxying input from the front end to the back end, so I'll just write your input, too." Boom. And there we go.
Now, here's the question. Is it going to allow that? That is the question.
So, it already got an idle timer that was fired because one of the workers idled out.
So, we're testing in real time here. Oh, we've got two that are waiting for permissions.
Look, it's paused on the Laravel boost prompt. Not finished. I'm approving.
Hell yeah.
So, it approved it.
And it approved it. Man, I want to say the f word. Hell yeah. That's freaking awesome. There you go.
H man. I So, uh, Hive Lord says, "North Texas weather is crazy.
Today's my uh today's my kids' birthday, my older kids. And so I went to their school to have lunch with them. And when you go to their school to have lunch, they put you outside cuz they just like in this era they just don't let people in the school, which I'm actually that's fine. I'm on board with that. Um so we checked the kids out. We ate outside and I you know I was wearing a sweatshirt but I had shorts on. I was freezing.
Kids didn't care because they had Chick-fil-A and they were thrilled to death to be out of school. But yeah, I was freezing out there. Also, their lunch starts at 10:50 somehow. So, it was like, you know, I got to work, you know, pushed a button and then had to go to lunch. Um, but it was awesome.
You know, Derek, as a fellow believer in the Lord Jesus Christ, sometimes I will drop the fbomb, too, when things are totally awesome.
Um, but I try not to do it on stream.
But yeah, when I'm off stream and something goes really well, let her rip, baby.
All right.
All right. You can use the solo MCP to uh spawn agents, as many as can work in parallel. That's up to you uh to start working through these to-dos. You're the orchestrator. don't write any code, but you can set solo timers to wake up when they go idle. You can see I'm repeating the same thing over and over and over.
And so that's where skills would come in. But like I said, uh like I said earlier, I am not using skills because I want to test the raw MCP versus, oh, the MCP keeps, you know, doing this stupid thing. Let me write a skill to to patch it. I as the creator want to fix the MCP whereas if I was just a consumer I would write a skill an orchestration skill or something. So there's a little mismatch between how I use it and how I expect other people to use it.
Uh will we see Solo get an integrated browser? You know Jode, it's funny you should ask that because what do you do? Well, not much very well right now.
But uh yeah, opened on March 21st.
Um this is also an interesting one. What do you do? H what do you do? That's why I use extra high fast because I've got 1 million things to do. Um but this is the big one. This is this is the one that we need to get landed first. I have a little document here.
Document. It's a piece of paper.
Um, so let's see. Um, there's like a 0% chance you can read any of that.
But this is uh momento mori. You want to see my momento my? This is my note right here that says you are going to die.
That's my momento my um but this is my this is my workflow or rather my my blockers list. Um so to get the browser, the editor, the diff viewer and the split panes done, I need to get the design landed first. To get Linux done, I need to get Windows landed. And to get uh two other things that I can't talk about done, I need to get mobile landed. So, so many things to do.
And you know what? We're doing them.
That's what we're here to do is do things. All right. So, this orchestrator, I need to rename you.
Guys, can we just talk about Can we just talk about this command P pallet? Um, so I stole the command P hotkey uh from Ghosty. So, if you go to Ghosty and you do command P, you can like change the tab title, terminal title, you can do like quick actions and stuff.
I really liked command P. So, I took command P and I made it actions that are scoped to this particular process. So, you've got command K, which is like your whole global action set.
You've got command E, which is like jump. Uh this is a recencyordered history list that you can jump to. But command P is awesome because it operates only on this process or rather it's scoped down to this process specifically. And rename is always at the top. Uh so you can have rename, restart, stop. When there is a timer, you can manually fire it, which is kind of nice. You can pause it if you're like, "Oo, hang on. I want to talk to you about something." You can open it in the activity monitor, which is cool. Um, I don't know if y'all have seen the activity monitor.
Whoa, cool.
So, this is everything that Solo is responsible for all the way down the entire process tree. So, you'll see here's the boost MCP. I didn't spawn that, but one of the agents that I spawned spawned that. And so, I can see everything that's working here. Um, but the thing that I use command P for most often is rename. So, you're the orchestrator, cool.
Going forward, let's spawn uh codeex agents to do the implementation. I think I accidentally said AMP earlier, but let's use codeex.
I am not made of money. So I can't have AMP doing all the implementation.
Oh, sorry. Jetpack Joe. Who's number one now?
You can use a GPT sub with AMP. Is that real?
Only a developer would add a process list. Yeah, it's like so down here.
You know how your agents love to do stuff in the background and then like, oops, I forgot about it. Uh, this this is really helpful for seeing what the hell is my agent up to back there. Um, and so you can click on that and be like, "Oh, you're running the herd MCP. You're running this MCP tools Lah files. That's another MCP that I need to remove." And then it's running the solo MCP. Um, so if your agent gets out of hand, uh, you can kill it.
I have not done an, uh, an AI token tally. Um, right now I currently subscribe to two Codeex Max and one Claude Max. So I spend $600 a month on subscriptions and then, you know, onesie twoosies on AMP. Uh, so Jetpack Joe Jetpack Joe claims he claims settings.
Boy, do I not know how to use this.
Next page says controlV.
This looks like a bunch of permission stuff.
Boy, do I not know how to use this.
MCP servers just solo. Yeah, Joe, you're going to have to help me out there.
Oh, on the website.
Um, I did have a remove boost skill or I had a deboostify. Um, that's back when they shoved literally everything into the cloud MD. Boost is boost is great now. It's so much better. Um, so I don't worry about that so much anymore. All right, cool. We're back to using codeex.
All right, you can go ahead and push all that then.
Yeah, boost is good now.
I think in the early days we were all trying to figure stuff out. Um, including the boost team.
All right, let's do a fresh review on the canonical operations. Um, I want you to give me uh a list of what's canonicalized uh for the MCP.
Uh, tell me what's not yet canonicalized for the MCP and then tell me where HTTP and CLI surfaces are still thin.
Yeah, Joe, you better you better you better be telling the truth, Jetpack.
Um, let's just log in here.
How cool is that blur filter, huh?
All right. Uh, settings.
Okay, I don't think there's anything bad in here.
It shows my email, which if you send me an email, I'm not going to read it, so that's fine.
Um, okay. I don't think there's anything in here that I don't want y'all to see.
Just double checking. All right, Jetpack.
So, I'm on settings.
I don't see anywhere to connect chat GPT.
Oh, I don't have it. Oh, convenient, huh? Well, that's convenient.
Old old Aaron doesn't have it, huh, Joe?
H suspicious.
Suspicious.
Yeah. I don't I don't know if it's secure or not, but it's awesome, right? Isn't it awesome?
It's pretty cool.
All right, so we've got some QA agents going.
Okay, you guys are good. Team licensing schema, you're working on that.
It is awesome. See, that's all that matters. Hey, like the stream, huh? We got to get to 100. How many are we at?
Can I get a report?
Uh, how many How many likes are we at on the stream?
If it's under 25, I'm going to be embarrassed.
If it's over 30, I'm going to be pumped.
Yeah.
Here's to y'all.
Just a little Celsius to keep the vibe going, you know. 36. Not bad.
Those are rookie numbers, though. We got to We got to pump those rookie numbers.
All right. So, here is the review.
MCP is in good shape.
So, um, if you haven't been paying attention, what the hell are you doing?
Are you like working or something? Get real. If you haven't been paying attention, here's what I'm trying to do with Solo.
So, I have had um, Can you see that? Is that legible?
Why is it so small?
Um, what?
Uh, yeah. Zen and Celsius. It's real real good for you.
Uh, so here's what I've been working with on Solo. So, I've had an MCP.
All right, I'll allow it. If you've been busy using Solo, you don't have to pay attention. Uh, I've had an MCP and then I've had a very thin Let's just do this.
Uh, yoink. I've had a very thin HTTP server. Okay. Um, and the HTTP server was originally just built for Raycast, uh, so that you could just like open deep links into Solo.
And then people started wanting to control it. And I thought, boy, it would be cool if Solo was like the most controllable, configurable, automatable meta harness on the market. Also, by the way, I recorded like five YouTube videos this morning and I'm trying to claim the mantle of metah harness. So, if you ever on Twitter see somebody like talking about what is solo meta harness, that's what we're going for. Solo is a meta harness. We're going to take a little detour because this is a stream.
So, oh, you know what? I wonder um I wonder if I can show you I was going to try to show you the chalkboard out there, but I don't want to mess up the stream. Uh so what I've been working on what what I'm trying to claim is my space in the market, right?
Because I got to have a space in the market. The LLMs are down here and then you've got your models. Nope. You've got your harnesses up here, right? So you've got um this is Claude uh this is codeex this is AMP this is open code and this is PI sorry Gemini then you've got your LLMs down here this is uh let's go in order opus GPT uh any model any model any model this is this is the state of the market right there.
What do you recommend for managing tasks involve multiple agents? Um, a beads integration. No, there's firstparty to-dos. So, I would use scratch pads and to-dos, and I'll show you that in a second. Um, and then up here you've got your meta harness. So, this is solo.
This is what I'm trying to claim right here. Um, and what differentiates a meta harness from a harness is it allows you, the metah harness allows you to still use your favorite harnesses. So, I am not trying to recreate tool calling, system prompts, compaction, all of that. I am just trying to encompass all of this uh and give your regular harnesses superpowers.
So I want to in people's minds I want to own the space of metah harness a harness orchestrator. Um it could be that's one aspect of it. Uh warp cursor conductor. Yeah. So um you can do orchestration with a meta harness. So, warp uh I think does warp allow you to do does warp allow you to use the vanilla clawed binary and codeex binary. If so, it counts as a metal harness, but I'm going to try to own that before they do. Um, cursor would be its own harness. So, cursor is its own harness.
Cursor. And then what is theirs called?
Composer, I think, is their model.
And then what was the other? Conductor.
I think conductor would live at the metah harness layer, which would make sense to me because I feel like my biggest competitors are Warp and Conductor. Warp has raised 75 million.
Conductor's raised like 25 million. I've raised 0 million. Um, so I think Warp and Conductor probably live up here. But one thing that I think conductor does is it doesn't use like the claw binary. It might shell out using claude P. I don't really know. I'm trying not to use conductor or warp cuz I don't want to poison my mental model. Um, so this is kind of how I'm thinking of it.
I think T3 code is similar.
I think it's similar to conductor um in that it uses the regular see it gets tricky because like conductor puts the guey on top of cloud code and the reason I'm not sure it's a total meta harness is because conductor recently was routing all of your 5.5 requests to 5.4 for, which to me makes me think, well, you're not actually using codeex if they're doing some routing on the back end that they can screw up. You know what I mean? Whereas solo, you're you're just running codeex. So, there's no like I'm not in the middle of that whatsoever.
Um, so this is my mental model and YouTube videos will start coming out about solo being a meta harness and hopefully I can gain I can own that term kind of like a meta framework.
So what I've been trying to do is this is how it has worked historically.
This has not existed at all. Um, this has been super thin. Just a few commands. This has been really rich.
This has had a huge surface area. So, instead, what I'm doing is I'm creating I'm creating an ops layer.
And then each of these, I'm just going to delete that. Each of these basically provides a translation layer that sits on top of the ops layer. So here in the near future, MCP, HTT, HTTP, CLI, and uh surprise, JavaScript sandbox will all speak to how do I do arrows?
It'll all speak to this canonical ops layer.
And so no matter what surface you go through, you'll have access to all the same tooling. That's the plan.
Super set. Yeah, I forgot about superset.
Uh, it doesn't have a headless mode yet.
Um, but this is going to allow us to move forward to headless mode. Um, it is on my list to have headless mode. Uh, there are a couple infrastructure changes I'm still I'm still making to get to headless mode, but yes, it's coming.
Okay. So that this is the canonical that's what the word you keep seeing canonical ops layer.
And so what I where I started was moving all the MCP tools into the canonical ops layer and then making thin entry points for JavaScript, HTTP and CLI. Um but first moving everything moving all of the MCP actions to this ops layer. Does that make sense?
Thought you might want to know since the point of this is a sign up. I'm guessing so. So, let's make sure people can do that. Sorry. The point of what is a sign up? I missed that. I don't have very far scroll back on here. Uh um unfortunately, I would recommend using codeex over opus altogether. At this point, it switched for me with um Sorry, I need to cough. Hang on.
It switched for me with Opus 47 and Codeex 55 or GPT55.
I don't think Opus 47 is as good.
Unlock access on your site.
Yeah, probably. We're working on it.
John, keep them coming. Uh, yes, Spooky Tooth, we are um, we're working on that right now.
Thank you for the heads up. All right, canonical ops layer. MCP is in good shape for the durable domain. Those now route through the ops layer rather than living as one-off MCP logic. All right, we're getting close.
projects, processes, to-dos, scratch pads, key value, store, locks, timers.
Perfect. They all go MCP resources, read through ops. Perfect. MCP not yet canonicalized.
Oh, you're a Sonnet boy. I think Sonnet is probably good.
Um, look at this orchestrator just spawning all these agents. I need to tell it to close them when it's done.
All right. Remaining MCP gaps are runtime control plane. Yeah, we need to move all those.
HTTP is still thin. CLI is still thin.
And interestingly, the CLI is just um it's just a layer over the HTTP. So I don't have separate CLI um ingress points. It's just HTTP.
Okay, let's uh let's first clean up um and potentially archive any scratch pads that have been superseded and let's create a new fresh working document. Um you can capture all of this about HTTP and CLI but primarily this is a northstar document for getting all MCP operations canonicalized.
So you said we still have runtime session and control plane. Um let's canonicalize those in our next pass. So, first clean up, archive, clear out uh old scratch pads. You can roll forward anything that you think is relevant, but I want to start from a fresh uh a fresh working document and also audit the to-dos to make sure that we don't have any lingering or left open um for the canonical operations.
Uh, and then once you're done writing the fresh scratchpad for MCP canonicalization, turn that into a set of to-dos and then stop and we'll uh figure out where to go from there.
What is your opinion on MCP servers in general?
Do you think implementing an MCP server is better than a CLI plus skill? I don't. Um, codeex is more literal. I don't think an MCP is better than a CLI plus a skill.
Actually, I don't even know if you need a skill. If the CLI is good enough, um, I don't know.
So, I think MCP has some advantages, uh, especially around like O and governance and that kind of stuff. Um, and especially for non like non-terminal based stuff. So if you're asking like a, you know, a business person or whatever to use a CLI, that might be a little bit of friction, but if you can just click a button and add an MCP to like claw desktop or something, that makes it a lot easier. Um, but CLIs, they're great.
They're totally fine. Uh, usually just for developers though.
Dude, Prime's the man. I love Prime. I have a lot I have a lot of respect for the Prime Agen. You're welcome, Derek.
That's very nice of you to say.
Um, all right. So, it's going to archive the old documents. What do you need?
Attention. Navigation is paused on a scoped edit. You're still working.
It approved that prompt.
Um, recorded the wake up in Scratchpad.
So, this one is interesting because it is um it's like using the scratch pad as a working document. So, we're in faster.
If we go to faster, did you know that if you put this little this little token here? So, you see FAS and then I do the token and then I do S again, it skips over to scratchpads. So, the little token is actually a separator. So I could do FA token S and the S is going to match scratch pads instead of faster.
So it's a good way to drill down. Um all right. So if we look here, you'll see let's come all the way down here.
It's keeping track of this orchestrator is keeping track of everything that it's doing. So if this orchestrator dies for whatever reason or it con it um compacts its context, it can just come over here and read exactly what it's been doing.
So it's just keeping notes for itself, which I think is really nice.
NCPs are good when you want a tool that is deterministic. You code exactly and there's always uh yes, I agree. MCPS and CLI are good for determinism. Skills are non-deterministic. I agree.
I think that the question is always MCPs or CLIs. Um, and then you can have a skill with the MCP and a skill with the CLI, but the question is like, is CLI or MCP better? And the answer is no. They're they're both great.
Horses for courses, you know. Does anybody ever say that? Horses for courses.
Um, all right. Who's stuck? You're still working on Stripe.
This is the big one. You're still working. Um, you can go ahead and close any sub agents that you don't need anymore.
You're a Kiwi Strawberry Man. Um, this one is Wild Berry. I'm partial to Wild Berry and weirdly the orange one. Um, I cannot imagine getting one of the fizzree versions. That sounds horrifying to me.
Um, we're going to we're going to open another desktop here cuz we need to add just a few little bitty things.
Uh, can you pull from origin main?
Why did it do origin main as one word?
We need to add a few little nice nicities that people have been asking for.
Um, can you pull from or Yeah, I don't know. Like uh energy drink that's just like flat water sounds horrifying to me.
All right. So, watch this. I have a bunch of to-dos. Um, I have a bunch of to-dos over here in this project. Everything has a deep link. There's a deep link for everything. And if there's not a deep link, tell me because there should be a deep link. So, I just copied a deep link. Can you use the solo MCP to look through the to-dos in that project? And there's one for command H. Uh, I need you to transfer that to-do uh over to your project. You can use the solo MCP who am I to figure out where you are.
Um, okay. Do you have to find a mindset that made AI engineering feel less overwhelming? 100% you do. Um you uh the mindset that I have found to made to make AI uh engineering less overwhelming is um is that I like code architecture and I like building things. I have also for 25 years, how old am I? 37.
25 years I've liked programming with my hands um since I was like 11 or 12. Um but I still love the architecture and the actual like producing artifacts and so AI just helps me do that part faster.
I still get to construct uh artififices in the sky in terms of architecture. I just don't have to use my grubby little fingers to write it.
Uh, when would I recommend Codeex over Opus? Right now, always. Unless, as somebody mentioned earlier, you want some creativity or some weird like personality, then go with Opus.
You're 37. Hey, you, me, and Taylor Swift.
I'm using Monologue as my dictation. Uh, Jetack. Yes, I need to be able to rename rename a project via MCP. Yes.
Uh, okay. I need you to go ahead and implement that. We need to make command hide the window just like every other Mac app. And let's also make it work on Windows, too, if we can.
All right, Josh. That's what I like to hear.
So, Derek, are you 1989?
Um, all right. So, I'm going to rename you to hide window.
All right, let's get Jetpack Joe in here.
Can we also add um an operation whereby users can rename a project uh via MCP but then later of course HTTP CLI etc. Jetpack how old are you?
80s all the way. Barely 80s kid 15 hours in solo. Woo doggy. We might have product market fit, my friends. Um I don't I don't mind having four checkouts. I really don't. Um, with Rust it can get annoying cuz the the dependencies are actually quite large. Um, and they can end up taking hundreds of gigabytes. Uh, but I just prefer it to work trees personally.
All right, Jetack add MCP project rename through canonical project ops.
All right, so we're we're in good shape.
Let's see what this scratch pad says.
MCP canonicalization.
This is a fresh working document. That's right. Superseded scratch pads. Look at that. See, we're just building this whole like this whole web of knowledge.
John turned 50 last year. Happy birthday last year. Um 1994.
You have an old soul that ranges from the 20s to 89. Um, current state, Northstar. I love these agents love to talk about North Stars. So, you just, you know, I just speak their language.
And I think where's project rename? Can I do command F on this page? Oh, that's a miss. I need to add command F to this page.
Move one narrow behavior at a time.
Add red tests before extracting high-risk runtime. God, I just love codeex, man. I just freaking love it.
They also love saying smoke test. They love a smoke test.
Boy, they love a smoke test. I think I have hundreds of smoke tests in this repo.
Where is Jetpacks added scope? There we go. MCP project rename.
Cool. All right. So, all right. You can start spawning uh codeex and codeex W agents. Those are both uh those are both actually codecs.
I just have two accounts to break up rate limits. So you can alternate between codeex and codeex w. You're the orchestrator. Uh so same deal. Don't write any code. Make sure everybody's working. Approve prompts yada yada yada.
You hit them with the yada yada yada to let them know you're still in charge, you know.
I'm building a maths tutor. You must be from the old world saying maths. I'm building a maths tutor backed by AI. My app has the syllabus and the database. Tracks progress. The AI does the tutoring. Oh, that's cool. I think there's going to be a lot of personalized um personalized learning moving forward.
Here come the codeexes.
See this is in in my not so humble opinion this is what good architecture thinking can give you. Um we have like we have a really robust architecture and we have good plans and so now all these codeex agents have to do is kind of just slap. They just kind of have to they just kind of have to do the thing. Um, so I spend a lot of my time just thinking about the architecture. Um, and it's this weird this is this is new to me. You can kind of feel like historically when we wrote code, you would know, oh, this is like this is brittle. This is shaky. But now you just kind of feel it as you're like talking to the agents. You're like, wait, hang on. Why is that like that?
No, no, no. this should be encapsulated or this should be a state machine or this should be a ring buffer or whatever. Um, and you just kind of like start to feel it in your bones.
Um, I don't know about the codeex extra token window.
I'm not sure. I've definitely yesterday I hit my 5 hour limit on codeex for sure. I had to had to cool off. Um, okay. So, how are you doing the command H?
Yeah, UK is the old world.
Okay, so this is the command H guy. How are you doing this?
Native menu items are also wired in Rust. Yes.
Correct. Hotkeys are tough in here because like depending on if you're in the terminal or you're on the sidebar or you're in settings, we have to figure out where to route it.
And this is our um talking about structure. This is our structure for hotkeys. Everyone has an ID and then uh a default combo which is a function primary because on crossplatform you have to handle different key presses. So we have that little helper function there.
See what the hell are you doing? Is Mac platform.
Oh, see this is this is just stupid.
uh you added a helper for is Mac platform and you're testing the uh you're testing the browser navigator.
There should be a first-party way in tari to know whether this is uh Mac or Windows or whatever. Honestly, you could look at the primary key combo function because I'm sure that we're differentiating in there. But we definitely don't need that helper that you just added. So, take another look at that.
Good job, Sally the goose, helping out the chat.
We love that. We love friends helping friends.
All right, let's see if this concedes the point that it shouldn't have done that.
I bet it's going to I bet it's going to roll over almost immediately.
If you have not liked the stream on YouTube, do me a solid.
be a bud. Agreed. I'm removing the navigatorbased helper, switching to the runtime branch to Tari's platform API.
Duh.
Come on, man. All right. So, what are you doing? You're doing.
You can tell what it's doing cuz it locked this to-do. So, I know that this agent is working on QA967.
And I could click on it. 45. All right.
Hey, we're cooking. We're up from 36. At least we're not flat. That's good.
That's good. That's good.
Did you all know about command shift E in solo?
So command E is your like jump pallet, but if you do command shift E, it only shows you the ones that have rung bells.
So if you hear, what an idiot. If you hear a bell being rung, you can hit command shift E. Boom. There we go. I split the ownership to avoid conflicts.
Yeah, you did, my dude. This is why I think agent orchestration platforms are not long for this world because the agents themselves are just good enough to do it. That's one of the videos I have coming out is you don't need a complicated like orchestration platform.
You just need to give your agents a couple of tools and they'll figure it out.
Um, Tari does have some update logic built in. I ended up uh not fully rolling it myself, but partially rolling it myself because let's um let's go to so this is the back end of solotterm.com and I ended up partially rolling updates myself so that I could do stuff like this where I roll it out slowly like this one got rolled out to 50% and I realized that's not that good. Um, and then this one made it to draft and I realized it's not that good. Um, so I have a whole platform on the back end that manages and this is how I do like beta testers and stuff. Um, so I host the manifest on solo terms API and then R2 holds the actual updates. But for like the download and update thing, um, for the download and update that is Tari, but I host the manifest and stuff.
Um, hey Loanski, are you going to implement a key map, my brother?
Boom.
You can change them all.
Anything you want, you can change.
Done.
No, the beta pool's never capped.
The beta pool is basically just people that are nice to me and can tell me if something's gone wrong before it goes out to the wide scary world. I removed the new navigator helper from use hotkeys.
Switch runtime to first party OS API obviously. Um, okay. So, what did you actually do?
There you go. Just import. There you go.
All right, that looks good to me.
I've had I've found myself No, not Windows yet. Soon though, having more fun in Swift versus the web. The native feel is so nice. Um, any top of mind for creating a web limited companion app based on this swift native app?
Um, I don't think there's anything special there to be honest with you. Um, I think it's basically just you do have to like for solo I'm trying to get certain primitives in place such that it can be controlled from the web but if it's just like a simple little companion you can have basically just APIs that should handle it. Um, yeah I don't think claude doubled its usage limits. I think it doubled its 5hour cap, but like I think your weekly plan is still your weekly plan.
You're still cooking on team licenses.
Uh, command H. Okay. Okay. Can you uh update the change log and then run pre- push?
Pre- push is my whole quality gate. Type check. cargo lint. It just does it all.
Oh, there was some other nice feature that I needed to add. Um, let's try to add command F to scratch pads.
All right, we're going to rename this command F.
Can you do uh Can you do a quick Well, it can be slow. I don't care. Can you do an investigation into the scratchpad detail view and figure out what it would take to add a command F to that view so that people could search for uh people could search for things inside of very large scratch pads.
So am I going to be out of usage? Uh this my work account is already in deficit. So, if I keep using it at this, it'll run out in a day. My regular account is okay, but then obviously when I switch when I stop using my work account cuz it runs out, I'll probably hit the limits. So, might be time for a third codeex.
Yeah, I know. I just I just talked to the agent. Just whatever's on my mind.
All right. What are you doing?
I kind of want to see how this uh redesign is going. But look at this. I'm not starting 995 yet because the dashboard worker is working in that area. Yeah, good. You figured it out. Good job.
Uh that usage app is an open-source app called Codeexbar by the one and only OpenClaw Guy.
So it's just a free open- source app.
Somebody mentioned another one on the stream the other day. Um, I forget what it's called.
I have subscriptions that cost $600 a month. Um, so I have two codecs and a Claude.
Yep. Mobile is cooking. Any interest in running OSS models?
Not really. Um, so part of how I think about this is I am currently I'm currently in an allout sprint to try to win. Um, cuz I'm competing against a lot of wellunded players and I want to use I want to use the best tools to try to do that. Um, so that's why I'll sometimes use fast mode. That's why I have two CEX accounts because like it is freaking it's now or never. Um, so that's how I think about it.
Yeah. Why didn't you do that? Yeah, keep going. Don't stop. Yeah, keep going.
Dispatch as many codeexes as you need um until we cross off all of those to-dos.
Oneman army. Me and my tokens.
Yes, use the limits as long as they last. I don't know that this I don't know that this is going to last forever.
Hello, Ino. Um, we're building a couple things. One is we're working on Solo, which is this app that you're looking at. The other is we're redesigning faster.dev.
And then we're also working in the solo term.com repo to add team licensing. So, I've had a couple of teams reach out and say like, "Hey, we want to, you know, centralize billing and stuff." And so, we're uh we're doing team licensing over here.
Brother, don't I know it, 600 is still way cheaper. Um, I'm hiring my first full-time employee sometime in the next like two weeks, I think, is when she starts. And boy, that ain't cheap. Hey, Rafal.
Hey, you new friends, if you haven't liked the stream, I'm struggling. I'm dying out here. Hit the like button on the stream.
Look at Look who it is. Awesome. Cling.
Yes. Just wait. Just wait for your tokens. Enjoy someone else burning tokens while your uh usage resets.
So, you're running pre- push. You're investigating command F. Let's see what you're up to.
Um, Derek, you're killing me. You're going to see it. Uh, the redesign is coming. Um, so what role am I hiring for? Um, I am hiring basically like a director, director of operations basically. Um, so I'm hiring someone to do everything except what I can do. So she'll handle emails, contracts, taxes, credit card points, booking, travel, first line, uh, response, sending out the newsletter.
Um, so I'm hiring somebody. I'm hiring somebody fulltime, and it's I think I think it's going to be awesome.
Um, hey, you already use solo. That's awesome.
Command F. What would it take to do command F? Okay. Medium frontend change.
No Rust or DB. That's awesome.
Scratchpad owns draft content. The plate editor ref. That's good.
Plate is also very good, y'all. Um, no, I'm not points maxing. Uh that is Kelsey. Um but this is this is breaking news. Um Kelsey will be moving on. I can't say where cuz that's her business, but she's gotten just an absolutely incredible opportunity. And so um she's going to move on to something full-time that is way better than anything I could offer.
So I'm super happy for her.
Uh, are you points maxing? No, but I need to be points maxing. I still use a debit card like an idiot.
All right, command F already. So, I'm trying to add command F to scratch pads because you'll get these these big scratch pads here. And it's nice that we have like the mini map on the side, but there's no command F, which is insane.
So, that's what I was asking codeex about. So, they say it's easy. Search open terminal only. the handler bails unless a terminal is visible.
Uh, if y'all haven't looked at plate as the markdown, plate MD, this is a very good Oh, that's not going to get me anywhere. Plate markdown. This is a very good markdown editor. This is what I'm using inside of uh Solo. Another reason I went with uh React is because stuff like this exists. They might have a view version. I'm not sure.
Um, I don't have any plans to add Hermes and Open Claw. I've never used either of them. Um, but if they have CLIs, you can just run them inside here already. Yeah, we were just talking about Discord notifications. I need to hook that up somehow. I'm still a Discord noob.
Uh, markdown is already optimized. Yeah, this was something I had to do.
This is the perks of using your own tool is I would load these markdown documents that were like, you know, uh many many thousands of characters and the whole thing would freeze and so I had to optimize it with chunked rendering.
All right, I would build Well, good, cuz you're about to build it. A scratchpad search state scratch open close query. Yes. Route search.open to scratchpad when selection kind is scratchpad. Correct. custom key bindings still work. Correct.
Um, add a small editor text search that walks the slate text nodes and builds ranges.
Um, correct. However, okay, that mostly sounds right. I do want you to take a look at the plate docs and make sure um and maybe the git uh GitHub repo. make sure there's no like firstparty search affordances because I don't want to reinvent stuff if we don't have to.
Um, but everything else about the hotkeys and all of that sounds correct.
But that's the one place I would push on.
Yeah, if I didn't build solo with solo, it wouldn't be half as good as it is now. And I think it's good now. Um, you know, I'm not reviewing changes as much as I used to. I'll be honest with you.
Hide window updated change log pre- push past. Okay. Okay. You can go ahead and commit and push. Um, yeah, I don't think the other agent has done any work yet.
Oh, I'm so glad to hear that. Thanks. Is it Vance or Vance? I don't remember.
Google Chrome quit unexpectedly.
There's not one. You know, you don't see that very often. Glad it wasn't solo.
Hey, even trillion dollar companies can't ship working software. That makes me feel better. The redesign is in full swing over here. Vince, I knew it was Vince Vance. One letter.
Redesign is in full swing.
All right. Who Who needs my attention?
Hide window. Push completed.
Okay, cool.
Um, awesome. Thank you. Can you look through the to-dos in that desktop one project and see if there's any other lowhanging fruit?
Solo was prevented from modifying apps on your Mac. What's interesting is that means some agent somewhere was trying to do something, but it still comes through Mac OS as uh solo.
How do I make sure things aren't broken?
Um, one, I test them all um quite a bit.
two, I have a lot of tests and I have a huge uh GitHub actions workflow.
Why is command shiftB not showing my There we go.
If we look at my GitHub actions workflow, let's look at one that passed.
This is my GitHub actions workflow.
So that's two jobs. That's 11. That's six. Uh that skipped cuz it didn't need it, but I think that's like 11 as well.
So I rely heavily on these tests.
Did building the OG PHP uh solo help you? Yeah, big time. I learned so much about uh OSC and ANC and all of that stuff. Um when I was building the original one by hand with my own two hands uphill both ways. Um PTY handles the whole like I learned just a massive amount and it was so much fun.
Um solo MCP really unlocked it. Yeah, that was the sleeper hit. I didn't, you know, I didn't expect that to be such a win.
Google's done better than Microsoft.
Poor GitHub.
All right. What did you find?
I don't even know what that means. Stop echoing harness bearer tokens.
Oh, that's a internal thing.
That's internal.
That's internal.
That's all internal boring stuff.
Yeah. 537 is done. So, you can you can mark that as complete. Uh, all the ones you picked are kind of internal and boring. Are there any product facing to-dos that we've logged that haven't been done yet?
My experience is that when I vibe code the app, vibe code the apps tends to broke sometimes in unexpected ways that are unrelated to new features. Yeah.
Yeah. That's when you got to start worrying about is your architecture sound and do you have uh excuse me, I got a cough.
Do you have the tests that lock in the uh the behavior that you're looking for?
These agents will like they'll overwrite tests. They'll write too many tests.
They'll sometimes write spellch check tests, which is like assert that the homepage contains the word docs. And it's like, is that really that helpful? But honestly, I would way rather have more tests, even if some of them are useless, than fewer. Um, so every time, you know, every time something breaks in a new and novel way, you either have to reassess, is my architecture wrong, like do am I missing a state machine if I added this if statement or this new status? Is it uh is there an invariant or something? Um, and then do my tests lock in what I'm looking for?
Yeah, talking out loud in an office would be kind of a bummer.
All right, you were right to push on that. I love being right.
God, being right. Being right is my favorite thing in the world.
It just happens so rarely, but I love being right. All right, the docs describe it as a search highlight.
Come on, man. And I knew we didn't have to build all that.
Come on, man.
Love being right. I freaking love it.
Nice. I love being right. Is there um can we reuse the uh command F chrome that we have in terminals as our like find you know text input next previous selector and then leverage the plate uh first party package uh to keep any new code that we have to write quite small.
Who said that? Everything I say is a fact. Well, all my opinions, but most are facts. Who said that? Was that me?
It sounds like something I would say.
Sounds like something I would say to Ian when I'm telling him he's wrong about everything, which I try to do at least once per episode.
Yeah, you check on that. All right. Add a oneshot notify when idle context menu.
It's not that default sidebar sort. We already have that. You can close 536. It's done.
You can close 736. I just completed that uh and pushed it earlier.
You can close 192. I think we've done most of that.
Nothing good there. Token Town is on the calendar.
Yeah, that does sound like me. Derek expecting her baby in two months.
Uh, here's what I'll say about vibe coding with a newborn. Newborns sleep a lot, but you should also sleep. Uh, when they're asleep, you should be asleep.
So, just make sure you're not killing yourself.
Oh, expose archive scratch pads to MCP.
I didn't realize that wasn't done.
Um, number 938 is interesting. I don't know.
I don't know how useful it'll be, but number 938 says add notify user to an MCP tool. Um, so I think it would be cool if agents could just fire off notifications that show up in Solo. Um, because you know, right now it's it's up to like the harness to decide when they fire those.
Um, I forget what number they are, but there's like OSC10, maybe that's titles.
Um, but any command can fire a notification, but I think it would be cool if agents had like a dedicated tool to fire notifications. That'd be awesome, huh? Any Easter eggs hidden in Solo? There are a few things that are feature flagged. 58 likes. Hell yeah.
Go, baby. Go, go, go. Um, there are a few things that are feature flagged, but nothing that's that interesting. Look, this is my Mac Studio, and it is set up as a GitHub runner. Um, and so it's currently running the end toend tests.
Isn't that cool? So like that GitHub action is running, but the actual compute is my Mac Studio and it's it's exercising all of these uh end toend tests that have been each one encodes some failure mode uh from some point in history. Isn't that neat?
That way I don't have to pay GitHub for their Mac uh their Mac runners. Sounds like a meta harness to me. Sounds like it. That's what I'm saying.
Um All right. Command F. What did you find?
Yes, the right shape is extract terminal search bar into a presentational find layer. Hell yeah. We love extracting stuff into a presentational layer. Tiny adapters. We love tiny adapters.
That's fine.
I'm okay with that. Uh, that sounds good. Write a full plan. um write a full plan for implementing this as a uh solo scratchpad.
And then before I have it start working on that, I'm going to have somebody else do that.
The runner can be local. Mac, help me understand what that question means. The runner can be local. Oh. Oh. Oh. Oh. The GitHub runner. Yeah, you can set up you can set up um self-hosted like GitHub nodes or GitHub actions nodes and they make it very easy. And so um the web still does all the orchestration but then the compute happens on your local machine which is good because I've spent like $500 on GitHub actions.
Yeah. So, an MCP tool that lets agents save knowledge is 100% on my list. Um, and there's also some clever stuff I think we could do around cuz I cuz I own the screen, right? I I own the terminal grid like I can I can read it. Um, I think there's interesting stuff we can do around like not even making it an MCP tool, but like having an agent output, you know, in in brackets like remember or learning or something like that and then have Solo pick up that that learning, you know what I mean?
You have to expose that machine to the web. Um, no, it runs an outbound poll.
So, GitHub has this thing where uh you basically just download it and then it you link it to github.com and then it runs an outbound poll I think looking for or maybe it runs its own local secure tunnel. I don't know. Um but I don't have to expose it.
Any app, is there any standard rule of thumb you're very opinionated about and wouldn't Okay, let's try that again. For any app, is there any standard rule of thumb you're very opinionated about and many and AI probably wouldn't think to think about to implement?
Um, yeah, I know my I actually created a whole new set of notification sounds um using 11 Labs. I'll show you in a second. I'm very opinionated about a lot of things, about anything, about everything. Um, I really, really, really want, as you can see with this command F example, I'm really pushing the agents to do less where possible. And so, like this one, this one started with the agent basically re-implementing the same commands for each surface, and I had to rearchitect. I had to stop it and be like, "No, brother. We just need one canonical layer." And then we can have thin adapters on top. And so I'm super opinionated and that's how this has grown. I mean, this is 400 500,000 lines of code, but I still feel like I have a good sense of where things are. Not like actually in the codebase. If I I've never opened this in my editor, but I have a good sense of like um I have a good sense of like what the architecture of the system is such that I can say let's not do that. Let's do this instead.
pricing changes to solo to make it viable as a business in the long term. I think we're getting close to viable as a business now. Um, and I am trying to win the distribution game. So, I'm trying to have it all. Um, if I were if I got money, I'd probably make it free.
All right. So, let's see what this scratch pad says. Command F.
Oh, see this is interesting. Look at this. I didn't know I wasn't on plate 53. I'm on plate 5211.
So maybe I need to update to plate 53 first.
All right. So, what I'm going to do is Can you read the change log for plate 53 and let me know uh what has changed and if it's worth upgrading while we're doing this.
Didn't know that.
All right. I found a small one that I want to do. Um, right now when you rightclick on when you right click on a project header or when you right click on a process, you get an add uh submen. I'd like to add uh to-do and scratchpad to that add submen. So users can click rightclick, go to add, and then do add scratchpad add to-do. And you can route that through the same handler that the uh command K pallet uses. So this is just another surface area for users to add to-dos and scratch pads.
Uh two steps ahead for usability 100% and constantly refactoring. Constantly refactoring.
Um open it in your IDE if you want a good scare. I would not dare. Uh, am I planning to raise the VC round? We'll see. We'll see what doors open.
Hey, Morgan. It's going great. Have you liked the stream, Morgan? Have you?
Because that's what it needs to go even better. Uh, we haven't yet escaped the underclass, but boy, we're getting close.
All right, so what do you think about plate 53?
You are still cooking on Canonical Ops.
Boy, you are just going ham on the redesign. That's good.
And you're still looking at the team licensing.
Okay.
Command pallet already creates scratch pads through that. Yeah, I knew that. I knew that. I freaking knew that.
So I this is this is the key right here.
Hey Derek, this is the key uh to add those same calls without introducing a separate code path. You have to have a mental model of what your code paths are because if you don't, it's just going to introduce new code paths. So I told it look at how the command pallet does it because we already have that somewhere.
You are one of the first. Good for you.
I appreciate that, Morgan. What a what a guy. I assume maybe. What a girl. I think I think Morgan is a guy. What a pal.
Um, okay.
You're still doing final tests. You're doing doc copy.
I would not upgrade to plate 53.
Okay.
All right. That's fine.
All right. I believe you. All right. Um, you can start dispatching codeex agents via the solo MCP to uh start implementing the command F and dispatch in parallel where possible, but make sure nobody's working in the same parts of the codebase.
This timer. This guy has just been cooking. He just keeps resetting his own timers. He's got two timers set. What do you need two timers for? What do you need that for?
Ah, not my problem.
Brilliant and beautiful. A man and his agents.
Yep. You're still doing orchestrator should handle that. But I'm just going to go ahead and approve.
Yeah, that's fine. Let's just get cooking.
Whoops. I just approved all those for you. My bad, dog. All right, hide window. You're actually now you're not hide window anymore. What are you? Uh, right click. bad.
Can I get a Can I get a check on the like count? How we doing out there? We doing great.
Bonjour, madame.
Even Even with uh the name Aaron spelled AAA, uh I still get a lot of mamms and misses.
62. Hell yeah.
So, if we're going for 100 and we have 62, that means based on my math, we're 62% of the way there. Huh. Interesting.
Cloud code now supports split views. Oh, really? How do I invoke that?
Cloud code natively supporting split views. That's kind of interesting.
All right, here goes some orchestration.
Find Chrome. Find plumbing. Find wiring.
Yep. Cool. Have fun, guys.
You watch the video announcement. What are you building? You know, what a great question. What a great question. An opportunity to hawk my wares. Does anyone still say that? Does anyone still say hawk their wares? Hawk my wares.
Hawk my wares. to actively and publicly promote, advertise, or shout out to sell goods or services, often in a street vending context.
It's a great time to hawk my wares.
Thank you for the question. Who said that? Uh, Faris Sarx 5552.
Um, let me hawk my wares. Right now, we're building what you're looking at, which is this is solo. This is a meta harness. You see, I'm starting to use the right term. This is a meta harness that uh gives your coding agents superpowers and we're building it out right here on stream. We're also building out faster.dev's redesign.
We've got four agents cooking on that with one orchestrator and we're building out team licensing for solo.
So, thank you for letting me hawk my wares. Merch is coming, baby. I contracted with uh some good old boys out of Atlanta to do some uh do some merch.
Uh sales have been solid. Let's see. Let me pull up Stripe on this monitor and let's I'll be honest.
I don't look at Stripe. Um it's too emotionally It's too emotionally taxing for me.
I make Kelsey do it.
But let's see, just for you all, let's see what Stripe says.
Uh, Slingshot is the name of the um Slingshot's the name of the merch company.
All right, this is this is fine. This is fine.
All right, you want to you guys want to see the Stripe dashboard?
You want to see some Stripe dashboard?
Make sure you make sure you hit the like button on YouTube. If you want to see the Stripe dashboard, hit the like button on YouTube.
When we get when we get to 75 likes, I will show you the Stripe dashboard.
How's that? 75. Somebody tell me when we get to 75 and I'll pull the Stripe dashboard over here.
Okay. You're still cooking. Monitoring two. Final.
Yep.
Canonical ops. Rightclick add. Command F.
Got to recruit a friend to like it.
What are we at? Still 63 or something?
Got to hit 75. There you go. It's working. I'm I'm figuring out how streamers do it, man. They just They just dangle. They dangle carrots. 75 likes. Is that real? 67. Oh.
All right. Keep keep uh keep liking it.
All right. So, um Arad, you were asking earlier, how do I make sure it works? So, this claims, yeah, but if I had emotes and Oh, hey Tilly.
That was a That's a great Laravel content creator right there. Um, Arad was asking, "How do I make sure that stuff works?" So, this claims uh this claims that it added rightclick to add. So, now what I'm going to do is I'm going to open this commands right here. I'm going to run tar fresh which runs the app in an isolated database.
We're at 70. We're at 70 75 likes. And I'll show you the dashboard uh the stripe dashboard.
So I'm going to start tar which is going to take a thousand years because it's got to build all of those rest dependencies.
And then I'm just going to go check and see if it works. Uh that's it.
Tilly, what are you working on these days? 72. We're getting close if y'all don't if y'all don't Hey, Tilly, see if you can drop your YouTube uh link or handle in here. I don't know if it'll reject you or not. Um but y'all should follow Tilly. Um yep, using one branch. This these are four different checkouts, but it's all one branch.
I do use separate branches for big stuff like the redesign and windows.
Um, but no work trees.
I'm too smallrained to use work trees.
She's on the native PHP train. Nice.
Native PHP boys are cooking. They applied to YC.
I told you this takes a thousand years.
It's really annoying.
All right. So, you passed your required QA checks.
Now, I run multiple checkouts at the same time. They're just in different directories.
It's just like they're all in different directories and they all point to uh they all point to origin main.
Vince, what are you cooking on native PHP? You've got a few that are like deployed to the app store.
Look at all these logs. Wake up.
I'll be honest, I'm pretty proud that I built a terminal that can handle logs like this. Pretty pumped about that.
Each checkout um has aenv.local local that declares um its database its SQLite database location and its vit port so that I don't have conflicts.
Same same kind of thing you'd have to do on work trees. Um it's just it's just not a work tree.
How's this redesign coming? I expect this redesign to look terrible when I launch it and then we'll spend a lot of time getting it all getting it all queued up.
All right. So, before we do any testing of the licensing, I'm just going to have AMP look over it. All right. Can you use um Well, you can cancel your outstanding timers. Yeah, whatever. I'll just cancel it myself.
cancel.
Um, okay. Can you spawn an AMP agent with the solo MCP and have it uh provide a review of the entire team licensing flow.
All right. So, we'll have it do that and then we'll look at it ourselves. 74. One more like and I'll show the Stripe dashboard. I'll drag it over here.
All right, here comes what is this?
Desktop 2.
Desktop 2 is starting. Now, in theory, I should be able to rightclick, add scratchpad to-do.
Do I like it up there? Oh, I don't think I like it up there.
Actually, I don't care. I don't feel like people use right click at all.
Cool. That worked.
Cool. That worked.
That worked.
That worked. Cool. 75.
All right, let's look at some money.
Money, money, money.
Um, okay. That worked. Great. Uh, you can go ahead and update the change log and then commit just your thin slice. Don't obliterate the other changes. Uh, but also don't commit them.
All right, here we go. You use right click. What? All right, who's ready to see the Stripe dashboard? Here we go.
So, Solo has made $26,000 so far. MRR is at almost 1,600.
So, it's working. I'm not rich. Um, but it's working. And more importantly than that, uh, a lot of people are using it.
So, I'm pretty pumped about that because it's a it has a pretty generous free tier as you might know as users. Um, and so until somebody hits more than four projects or more than 20 processes, they just they just use it like for free and that's the goal. That's how we get them, right? So, yeah, I'm pretty happy. Um, I think all of these refunds are fine.
That's like what, seven people, maybe eight people? No, seven. Got to be seven. Um, seven people that probably suffered through the point 0.5 series and I can't blame them. Um, yeah. Thanks y'all. I I feel I feel really good about it. I'm like kind of shocked, really happy, equal parts terrified. Um, cuz I feel like I'm on a runaway train. But you know what? The bridge isn't out up ahead. It's just moving really fast. So, I'm okay with that. Uh, but yeah, it's really nice. I'm I'm both proud of myself and and very happy. Um, I do have analytics running. I have Appase, which is like a privacy like a privacy first uh desktop analytics thing.
Hey, that's great. You hit the free tier in a few hours. That's awesome because I think it's generous, but I also feel like once you real like once you really get into it, you're like, "Oh, hell yeah. I got to I got to I got to launch, man. I got to launch all these agents um or add all these projects." And so, I feel like it is generous enough to get people using it, which is the goal.
Uh, hey, Vince tweeted about us. Thanks, Vince. That's nice.
Oh, yeah. That's 26. That's 26 topline.
I've spent a lot of money on website design. Will King design. Yeah. So, we're not like we're not rolling in the dough. Um, but we're laying the foundation for sure. Um, okay. KO, you can push that.
All right. What does AMP think about?
So, AMP has this 2y scroll back um that sometimes is hard for the other agents to read. So, you can see it just requested, hey, can you print that again for me? which is just if that's not AGI.
I mean, it's not, but it is pretty cool.
You can also tell it to write a scratch pad if you can't read it.
That's okay, Brad. I'll get I'll get work trees. I figured out um I've figured out how to do work trees in a way that I'm happy with.
Um and the canonical ops is part of getting there.
So I figured out a way. My problem with work trees, well one problem is I don't use work trees, so I don't have a lot of opinions on them.
Uh the other problem is everyone uses like work trees slightly differently partially because they they are a little bit cumbersome in my opinion because you have to set up like we were talking about earlier the ports and the database paths and all that kind of stuff. Um, but I figured out a way I figured out a way to get work trees copy on right or discrete checkouts. Like I figured out a way to satisfy everyone. Um, and it takes a little bit of plumbing like I'm having that's part of why I'm doing this canonical ops migration uh to get um this whole command surface area unified because to get work trees in I need to expose certain commands um to the runtime and then I think everyone's going to be thrilled to death with the work tree implementation. So, it'll be it'll be fun. What is uh what is conductor spotlight feature? I don't use conductor. What What is that? Uh copy on right. Yep, we'll get there. Smart suggestions.
I don't know if I understand the context of that question. Uh yeah, Morgan, if you don't know what they are, you don't need to know. I don't use them. The agents can use them.
I'm I'm fine if the agents use them, but I just I don't use them. Nate McGrady, Mr. X himself, formerly, I believe.
Nate, where are you these days? Are you still at X.com, the everything app?
All right. What did that see? Okay. AMP wrote the full review. So, I closed the agent.
Robin, Mr. Tailwind himself.
Um, okay. Robin says, "I use two work trees, one with the current PR and the other with the main. That way I can compare output and performance by running both executables." Oh, that's a Yeah, I imagine that's for the Tailwind compiler.
You can push spotlight on a work tree that gets rynced to your root dur so you don't need to run multiple local M's.
Bro, that's what I'm talking about.
What?
Yeah. So, the thing that I'm building, which I'm keeping close to the vest for now, will allow you to do that. Um, but it won't force you to do that. Smart suggestions point out using scratch pad making two using timers MCP. Um, that's interesting.
No, it doesn't have any of that. But it could, but I would need a model. I would need an intelligent model to watch the terminal grid to kind of make those suggestions. And I don't have I don't have any local ondevice models yet. Um, speaking.
Okay, here we go. I just got an email 1 hour ago. I just got an email. I sent an email late last night at uh what time did I send this? At 10:07. I sent this email from bed. Um 19th not 19th amendment just 19th.com.
So I saw this on Twitter. I'm sorry. I saw this on x.com the everything app. Uh it's called it's called 19th and um I saw somebody tweet it and they said that this is a way that you can train small Oh, cool. This is a way that you can train small specialized models. Okay.
Turn your examples into small fast models product tasks to beat frontier LLMs on accuracy, latency and cost.
Right. So, um, GPT5, Opus 47, they're good at everything.
GPT5 is better at everything, but they're good at everything.
That means they could be less good at very niche tasks, or they could be good at them, but they're still incredibly expensive. So, I saw 19th.com saying, "Hey, we can train a model that does one thing, and it would be tiny."
um and you could use it locally. And so I thought this is an idea I've had for a while, but I've not had what's the Iron Man thing like the technology of my time limits me or something. In this case, it was uh my brain limits me. I'm just I don't know. I'm a developer. I'm not like a model guy. Although I used to be a model when I was a kid. Um anyway, that's a different thing. So I saw this 19th.com make a model from your data in minutes. So, this is like a a new service that will uh build you a purpose-built model, which is like, oh, cool. Like a I could build a tic-tac-toe model.
All right. Cool. Like, why? I mean, I guess if you're building, you know, one player tic-tac-toe game, that's cool.
Um, and here's like their, you know, madeup chart that puts them in the good spot and everybody else in the bad spot.
Whatever. That's fine.
um 310 times cheaper.
So, I'm going to tell you where we're going to use this in solo, but just bear with me. We're going to scroll. They put their their design engineer put a lot of work into this. So, we're going to scroll through it. Ooh, beautiful.
This I am a sucker for how big was your context as a child? Is anything like my kids? absolutely tiny.
Okay, so I reached out to these folks last night and said, "This looks cool.
Here's my idea. Um, I would like I would like to train a very very very small model that can run on a user's device."
Okay, so on your MacBook, on your maybe Windows machine, who can say? um with the express purpose of figuring out if a terminal is waiting for user input.
That's it. Is it is it stopped at a permissions prompt? That's all that I'm asking. And so I sent him an email last night that was like, listen, I've got this app Solo. I've got um access to a virtualized representation of the grid, uh which is like the terminal. And so if I had a couple hundred maybe a few thousand examples of any coding agent that is stuck on a uh permissions prompt user request ask user tool. Could I train a very small model that then I could run on the user's local machine and then surface a notification if we get like a 95% confidence that a model is waiting on the user's attention which would be cool right cuz like right now most of what I have to do is curistics based um I have to like you know are there pty bytes you know are there bytes flowing through that means it's still working, but some some agents will still update the screen when they're waiting on a permissions prompt. So, it's not perfect, but if I could have an intelligent model that looks at it and says, "I'm 95% confident this is waiting on the user, that would be freaking awesome." And he just emailed me back.
What you're describing is actually a really great use case. Hell yeah.
Assuming you can extract the text from the virtualized screen, I can.
I would think a small binary scoring model that returns something like permission prompt confidence could be achieved with a sub one megabyte model.
Rough guess is a few thousand labeled examples of both positive and negative cases needed across different CLIs, package managers, etc. It feels totally possible.
It is so on sub one megabyte. That's freaking insane. You can just train models. Hey, where did you get that? Where did you get that catchphrase? Um, so there you go. That's going to be fun.
That's going to be real fun if I'm being honest. Um, all right. for licensing.
Yeah, I'm fine with that.
All right. Merge merged with follow-ups.
Um, yeah, that's fine.
Okay. Okay. Newly synced pools get a hidden claim token hash. The dashboard regenerate claim link even though no manager has ever seen a usable URL.
Yeah.
Oh, I need to cough again.
You know what? Also, I want a cookie and a Celsius.
How long have we been How many likes do we have? Can I mark off the 100 likes thing? Can I do that yet? Are we there yet?
Then I'm going to go get a cookie.
Um, you need to fix this.
Yeah. Why don't you um Why don't you add two red tests for the medium findings and then you can drive them green. Um the low findings I don't care about team licensing.
Uh this voice to text is called monologue.
Okay. You are still working on so even when a agent has a timer set you can sometimes sneak in there and talk to it anyway.
Um, how can you give me a status update on how close we are to the redesign being at least able to be looked at locally?
Canonical ops, baby.
All right. So, it's just waiting on Can you get a cookie? You can. You're a grown-up. You can do whatever you want.
It's your life.
All right. What does this say?
19th model would be local and solo. They could opt out if the user finds it not very accurate. Sure. Yeah.
Yeah. If they if the user finds it not very accurate, then I have failed. Um, but I I will add an opt out for sure.
That's fine. All right. So, you claim you added command F. Do we still have We don't. So, I'm going to have to launch this bad boy again. It's going to take a thousand years.
All right, I'm going to use that opportunity to go get a cookie and a Celsius. So, uh, don't go anywhere. No, I'm not doing any normal coding. Not even a line. I'll be right back.
84 likes. We're getting close.
All right. So, Solo 2 came up and it claims that uh this now has command F.
The link is sent via email. Yeah, we're going to have to see. Um I think my intent is that they uh will have to log in to get their license. So, um, it built whatever we planned. Now, we have to see if it's actually any good, and then we'll correct it. Command F does not belong there.
That's ugly. It's in the wrong container. All right, that's fine. Um, control. Oh, look at that. Oh. Oh. Oh, it just started selecting.
Okay, that's a huge problem.
Yeah, disaster.
Nevertheless, we press on. All right.
Uh, which one do I want to do first? Um, let's do this one first.
Oh, that's new. Oh, that's a problem.
Oh, that is a problem. Um, okay. I'll explain that in a second.
The uh Chrome is anchored to the wrong parent element. You'll see how it shows up here inside of the markdown. It needs to show up in the the top right of the pane.
So, that bug that you just saw, uh, Canban's done. I just don't know if I like it.
like it works great. I just don't know if I like it.
Uh the screenshot app is dope. Um and then I need to explain the bug that just surfaced.
Cleanshot X.
This the freaking best.
Highly recommend.
Um all right. Let me describe this bug that I just saw. So, the other day somebody asked, could you add drag and drop to the sidebar to add a project?
And I thought that's a great idea. So, I added drag and drop as a target for the sidebar. Um, but I don't think I scoped it down to just folders. And I don't know if that's possible. Uh, but I didn't. I definitely didn't clearly scope it to just folders. And so now when you drag an image, it says add project, but you're supposed to be able to hover over a process like this to select it. So you can, you know, select the right process and then move over here and drop it. That's the intended use case for single drag and drop. Um, so I got to figure out I got to figure that bug out.
Let's work through it together. Hey. Um, rename drag and drop.
All right. I need you to help me track down a bug. We recently added uh a drag and drop target to the sidebar with the theory that that would be for dragging and dropping projects in. But I don't think we differentiated projects versus individual files. So now when I drag an image over on the sidebar, one, it says add project, which is insane. And two, it shows the full uh it shows the full sidebar drop target even though what I'm really trying to do is drag over a single process to activate it such that I can drop the image in the terminal. Uh can you do some investigation and explain back to me the root cause of this bug? I don't want it to actually go fix it. I just want it to see if it can figure it out first. Doing some AI magic. You know, we're trying. Clean shot. Yes.
Tilly, what widgets are you working on?
All right. So, this one should be Well, you know, it moved it. It doesn't look great if I'm being honest.
All right. Good try, friend. Good try.
Swing and a miss if I'm being honest.
Um, okay. Okay, it is now positioned in the correct place, but we have some sort of like um overflow, some sort of DOM issue where it just looks insane. I don't know. You figure it out.
Are you doing like a whole Tilly, are you doing like a whole like um like a plug-in play like you're going to sell widgets and stuff for native PHP? Like a platform play?
Save you for now. Let me get out of the way.
I see it.
Um, the design of Solo as it exists is just guess and check. Just try to make it look a little bit better every day.
Nice, Tilly. I love it.
Oh, I see. Native like iOS widgets. I gota That's cool.
DN D root cause. Here we go.
There we go. That's all I needed to know. It does not check whether the path is a directory.
That's all I That's all I needed to know.
Is it possible to know that the item being dragged is a folder?
If so, that seems like a quick guard to add.
Okay. So, you think you fixed that? In fact, you did not fix that. Let's uh go out.
Maybe hot reload didn't get it.
Nope.
No, you didn't fix it. And I don't think that was the issue. Notice how the um notice how the Chrome that contains the search input uh is not growing to actually contain it. It looks all like shrunk down. I think something in Oh no, I ended DOM layout is uh escaping or not sizing properly. Why don't you start by describing the exact DOM layout to me and we'll see if we can catch it while you do that.
Sometimes if you ask it to uh explain it, it just figures it out while it's explaining it.
Can we start an agent in a project with a task on auto start of solo? Oh, that's interesting. No, not right now.
But that's interesting cuz that would be cool.
It's a good note.
Um, okay.
Hell yeah. Rust, baby.
I am the rubber duck.
See, it said, "Yep, I think the issue is the rapper." Yeah.
See, I don't even need to read that. It said I think I found it.
All right, go ahead and fix it.
Hey, thanks for hanging out, Joshua.
I have found monologue to be good.
All right, you pushed so I can kill you.
So long and thanks for all your good work. Um, okay, so this is saying Tari doesn't report, but Russ can cheaply stat them.
Cool.
Add a small tar command.
Classify cache on drop. Yep.
Yep.
Cool. All right, let's add that narrow fix.
If you tell it if you tell it to keep it narrow, sometimes it keeps it narrow sometimes.
Telly, where are you based?
There we go.
Um, if I wanted to call a specific skill, I'd probably just type the dollar sign or slash um and type the skill out myself. I didn't know you were in England.
I didn't know you were from Denmark.
All right, that fixed it. But I feel like that's still too wide.
Yep, that fixed it much better. There's still an odd gap uh between the text box and the controls.
I'm not sure if we're using like a fixed width or something. Uh but can we narrow that gap and or increase the size of the text box to fill the space?
How much did it take? Uh, how much of my life force? How much money? How much time? How much what? Alexander.
It took everything.
That's a reserve slot that's supposed to be there.
Are you hot reloading? You sure are.
I didn't realize that was supposed to be there.
How much energy?
All of it. I've slept in the office uh more times than I would like to admit.
Yeah, it takes a lot of work, man. Um yeah, go pick up your daughter. That's way better. Thanks for hanging.
Um, I probably started in December and I think I launched I think I launched in February maybe. Yeah, it took it took just a huge amount of of life out of me. Um, yeah. It's not easy for sure. Are you rebuilding? That's what's going on.
You're rebuilding. Okay. So, you're claiming that that spot is reserved.
Yeah, it sure is.
Um, okay. Well, it's still too big.
Okay. I didn't realize that was a uh a reserved spot. That's fine to leave the gap. Um, but even when there are matches, it's still a little spacious.
So maybe we can collapse that down just a bit.
A solo book. That would be awesome.
Yeah, we're getting browsers and diffs and we're getting it all.
Uh yeah, I started on I think I started solo on 54 or I'm sorry 45. Um and then 46 and it hasn't I haven't done much 47 on it. It's mostly been GPT.
Yeah, I still enjoy the journey. I mean, there's there's hope, right?
Like it's working. I've built so many things in my life that didn't work and it's pretty hopeless. um and depressing, but this is working and it's a lot more fun. Um still exhausting, which is why, you know, I still have four kids at home, so that's why I'm drinking a bunch of Celsius. Uh but yeah, it's more fun when it's working.
Sure, I can show you an orchestrator loop right now. So, this one is um this is just a codeex agent. There's nothing special about it. I tasked it uh we started with a scratchpad which is just like I talked to the agent. We came up with this big plan and I told it to write it down as markdown and it did because it's a machine and it does what I tell it to do. And then I told it to turn it into to-dos. Um, and then once we have this graph of to-dos built out, which I didn't do, the agent did all of that. It turned it into to-dos. Uh, then I told it, great, just start spawning agents.
And so this orchestrator has spawned this agent here. And you can see this agent has locked that to-do. And it's just it's just cooking on this plan that we talked about. You can see up here if I click on it, look, the orchestrator is keeping notes on it. And I I didn't I didn't tell it to keep notes. I just said start spawning agents to work on those to-dos.
Um and so it is it wrote all the to-dos.
It, you know, add comments as it goes.
Um, let's do all statuses, group by flat list, sort by created, newest, first.
So, this looks like a,00 1,1003, 10,000 10002. These are the to-dos that it's still it's still working through here.
So the basic theory is you either do deep research on the web and get this final markdown document and then you bring it in here too. Or you just talk to the agent and you're like, "Hey, I want to do this. Let's talk about it. Oh, no, that's a bad idea.
This is a good idea. What about that?
What about this?" And then finally it's like, "Here's what I would recommend doing." And then you just tell it, "Uh, write that down. Write that down somewhere." And then after that, you don't really have to worry about context compaction because it wrote down the Northstar document and then it's going to dispatch out to sub agents uh to do the actual work and they all have fresh context windows and if they want to read the Northstar document, they can go read it. So that's kind of the orchestration loop. I have a video coming. I recorded it this morning. I have a video coming uh soon.
GPT is my jam, too.
Hey, thank you. I'm glad you're enjoying it.
Uh, Jen, I don't know how to pronounce that. I'm sorry. Thanks, Jen. I'm just going to call you G I N. Jen, I'm glad you're enjoying it. That means a lot.
All right. It claims that this redesign is ready. I'm not ready to look at it yet. Um, okay. Cuz we still need to fix.
Yep.
Um, I need that.
Not that one. Not that one. Not that one. Not this one.
Okay, that looks good. Uh, the second issue is after I So, I do command F and I type the first key. So, I type the letter A or whatever. Um, and then plate highlights all the matches. One, there's a little bit of a layout shift in the highlight, but the more important thing is the uh focus switches to the plate area. And if I try to keep typing, I end up just replacing uh text that has been matched. So, you might need to look at the plate docs for that. Uh, unless you can trace it locally or we can add some logging to figure out why we're losing focus there.
What would you say is the ideal way to prevent a to-do from being modified by other agents? Agents can lock to-dos.
Um, and I believe that prevents I believe agents can still comment on them, but they can't change them.
Any skills you like working with? I haven't used grill me. I have an interview me skill. Um, but at this point, I don't know if I'm just like down on skills or what, but I end up just talking to the agent. I feel like the skills are maybe less and less necessary. I don't know if that's true, but I end up just talking to it.
Uh, how do you make sure you're not introducing regressions? Uh, a massive test suite, a lot of local testing, and what I hope are good agentic engineering practices.
There you go. Found the problem.
editor.tf.focus.
Oh, hey Dan. That's cool.
Wow, that rules. Um, some of them can get quite large. Uh, some of the to-do or some of the scratch pads can get quite large. Um, I Yes. So, I have a fix for that. Um it's not shipped yet, but I did just introduce um yeah when I enits I do I haven't enit like two months. Um so for scratchpad huge scratch pads I just introduced a uh replace section tool. So it already has appended which is nice because it can write without reading at all. Um, it ha, excuse me, it has read where you can read headers only. It has read section where you can read a section only, but it didn't have right section. So, I added right section. 87 likes. We got to get to 100, man. Um, don't even use claude MD. I know. I feel like I'm I feel like the models are just getting better at stuff. Um, so Dan, yes, that is a that is a paper cut that I ran into as well. Um, and so that'll be out in the next that'll be out in the next release.
Um, I am neither reviewing much of the code nor am I pure vibing. Um, so I know we've been going for many hours here, so I don't expect you to to watch it all, but there have been times when I've caught it doing something stupid and I'm like, "Hey, brother, let's let's not do that." Um, and I think a big part of not pure vibing is like understanding your architecture. So you can tell it we already do this.
Let's route it through that or let's extract this as a common component or like let's let's extract this as a set of commands like uh we started the stream talking about this.
So originally MCP was huge, HTTP was small and I have directed it to create a canonical ops layer and then have like a thin slice for each on top. So it's a lot about like uh understanding your architecture.
Basically the entire documentation for my OSS maintenance process.
Oh yeah, that would be quite large. Um yeah, that'll be that'll be helpful for you for sure. Um so sorry that doesn't exist already.
Um, but I did fortunately a while back introduce progressive rendering for Scratchpad. So, at least they should render without freezing your uh your WebKit thread.
All right. So, it claims it fixed the focus issue. So, let's do this. Let's come out here. Command F S O L O. And then if I hit enter, I mean, it works. It doesn't quite show me what the current um does it scroll like? What's a word that's way down here? Handoff is a word that's way down there. Handoff.
All right, we're getting closer. Um okay, that fixed that issue and we're getting much closer. We fixed the uh focus being stolen. We fixed the uh layout shift. The final thing is I can't tell uh I can't tell which I can't tell which match is the current match. So like if I match a word at the very top of the document and a word at the very bottom of the document and I hit enter and it goes to match two of two. First of all, it doesn't scroll the plate pane. Second of all, there's no visual indication as to which match is my, you know, my current match.
Uh, command F for scratch pads would be great. I'm adding it right literally right now. Um, yeah, the the stream title's a little bit tongue and cheek about, you know, people claim that AI is going to create a permanent underclass. I don't think that's actually true, but it's a good bit. How do you find AI is for tasks like database design? You got to talk to it. You got to talk to it. We earlier on the team's thing, we went uh we went through some database design stuff and it got it it got it pretty right. There were a few things that we had to tweak, we had to tune up. Um and it's it was pretty good, but um I still I just talked to it.
All right, drag and drop. Did you fix that?
All right, I added a classified drag path tar command that stats the paths and returns is directory.
All right, you think you fixed it, eh?
Let's see if you fixed it, my friend.
All right, so what we're going to do is we're going to come up here. We're going to get uh just any image whatsoever.
Yep, that's fine. Now, supposedly, hey, you did fix it. So, the issue was when I was dragging a single file, it was treating it like a folder. And when you drag a single file over the sidebar, it should allow you to highlight a process because like let's say you start dragging and you're like, ah, f, I'm on the wrong process. Well, you can highlight it. Boom. And then drop it over here. That's exactly what we want.
So the counterfactual to that would be if I go to solo term and I drag a folder add project.
Hell yeah. Yay.
Tada. All right.
Nice.
89 likes. Come on people, throw me a bone. We got to get to We got to get to 100. These are rookie numbers, but we got to get there. One more question.
When you do UI design, do you do wireframes, prototypes in Figma?
Neither. I don't do I just kind of yolo it. Um, and honestly, I end up hiring designer more often than not. Um, but once I have a designed system, I just basically force the agents to make good like React components and then try to just try to match them basically. Um, okay. Added classified drag. Yes. Yes.
Yes. Yes. That's good. Perfect. That worked. Great. Nice work. Can you please update the change log and then commit your narrow slice? Another agent is working on the scratchpad area, so don't obliterate their changes, but also don't commit them.
This orchestrator is still cooking.
So over in desktop one, we're working on this canonical ops migration so that we can have an HTTP interface, CLI interface, MCP, and a JavaScript sandbox for future.
Any suggestions on email receiving programmatically or as a human? Because I have suggestions for both probably.
Slice. They love to say slice. They love it.
Why don't you make a to-do list for that? This is I just asked you to like just do a little bit.
All right. You think team checkout and sync is done. Um confirm the new tests failed.
Did you drive those new test screen?
Get real. Did you drive those new tests green?
There we go.
Mhm. Mhm. Mhm.
Mhm. Mhm. Mhm.
I have opinions. Hang on.
I'm eating a cookie.
What new uh stripe plans do we need? And are those wired through the filament um admin area already?
You guys want to see the overhead shot?
Boom.
Pretty cool, huh? Now that kills the music. I forgot that that kills the music. My bad, dog.
All right, email.
Um, not is correct. Uh, Reand added email support recently.
Um, so I have many domains. I've got a lot of drinks. Too many drinks. Um, I've got faster.dev. I've got tryhardstudios.com. I've got soloterm.com all these all these do screencasting.com um for purely for humans um I set up domain I set up the domain on cloudflare and then I set up the MX records I set up Cloudflare email routing for inbound email and I point it to Fernand um this is our help desk.
This is an indie maker bootstrapper. Um, so you probably want to see that on the screen, not the back of my head. Um, this is our help desk. So all Tryhard Studios emails, uh, all of our various domains come into Fernand. So it goes to Cloudflare MX records and then it gets forwarded on to Fernand. Um, so we can receive and send via Fernand. We also have uh sending records set up on Bento.
We have sending records set up on Bento to send like mailing list and transactional emails.
Um, so that's how I do most that's how I do almost every domain. For Tryhard Studios, we have an actual G Suite account. Um, because Kelsey and I both have, you know, Aaron at and Kelsey at, but for like Solo, it's uh hello, [email protected].
That goes to Cloudflare to Fernand, and there is no Aaron atsolletterm.com.
for programmatic email receiving uh like if your application needs to receive email um resend has recently announced that Cloudflare has recently announced that so you can like route I don't know if Cloudflare recently announced it but Cloudflare has it you can route an email to a worker they want everything to go to workers um so you can route an email to a worker so if I were doing programmatic email receiving Um, if I was already on resend, I would use that.
Otherwise, I would use Cloudflare. Um, and I would probably, what I would probably do is I would receive the email at Cloudflare, route it to a worker, and then dump it to an R2 bucket, and then I would have my application side uh either get notifications or pull the bucket for new emails. So, I would dump the raw email body to an R2 bucket and then I would basically work it like a queue.
Um, Fernand is freaking awesome. So, the only thing with Fernand is um to my knowledge, I don't know if this is I don't you'll have to forgive me. I'm not sure I'm super up on MX Records. I think uh I think Fernand requires well obviously it requires MX records so I don't know if you can have like Fernand handle just like one or two email addresses or if it has to handle the whole domain. I don't really know. We just use it for the whole domain. So like all of solo faster screencasting the MX records just point at Fernand and it works great.
All right, drag and drop is fixed.
Committed.
All right, you can push that.
Uh, somebody do use pre-commit hooks.
CI/CD. Pre-commit hooks. No. CICD like you wouldn't believe.
Um, first of all, here's the CI.
So, you can see a bunch of those were skipped because it detects what files were changed and tries to save me money, which I super appreciate. Um, a huge CI workflow, but more importantly, let's find one that ran everything that didn't run Linux.
There we go. All right. So end to end it runs 12 end toend shards. It runs six Mac OS endto-end shards. It runs 14 Linux coverage jobs.
Um so this gives me great confidence. But more important than that is the only way that I can release solo literally the only way that I can do it is by coming to this release workflow and running the workflow. What that means is I'm never going to mess it up, right? I'm never going to like be super tired or it's going to be super late and I release a bad version or something. I mean, I could tag the wrong version. But what I can't do is mess up the deployment pipeline. The deployment pipeline does not, for goodness sakes, live on my computer. Can you imagine?
And so what this does is this has a few smoke tests in it. First is it makes sure that CI passed. That's the first one. If it hasn't pass CI, sorry brother, you can't release it. Um and then it builds it. And then it has two smoke tests that just make sure uh that it can launch on Mac OS cuz it's only Mac for now. So if it like immediately crashes, I'll know. Hey Jordan. Um and then it publishes it which pushes the artifact. Well, it actually pings my backend. Uh and then filament or uh Laravel pulls it in and then I have a filament panel to do the actual release.
Uh one thing that like uh one thing that I'm always trying to do is make things um deterministic and not rely on me.
I just don't I just don't want to have to remember, oh, how do I how do I publish this? How do I publish this app?
Or oh, shoot, like I spilled coffee on my computer that had the scripts to or the keys to publish. No, dude, get real.
Oh, it took me forever to build out this whole workflow. This is the accumulation of uh 2100 commits um forever.
Never going to give it up. Never going to mess it up, God willing.
All right. Who's doing what?
You claim that you fixed I don't know where that text is coming from. I don't know if that's me or if that's codeex.
Um, you claim that you fixed this. So, if I go here and I search for rewrite, which is at the bottom of the screen, command F, re Hey, it scrolled down.
That's good. That's good. Okay. Now, if I do open, and I hit enter. Hey, look at that. We have a little uh I call that good. Hell yeah.
That's a win.
We got command F, baby.
Look at that, my friends.
Um, desktop rendering on Wayland, brother. I don't even know what that means. Um, so I'll need to talk to you when it's Linux time.
All right.
Command F. Done.
Yep, that looks great. Nicely done. Um, you can update the change log, run pre- push, and then commit after it goes green.
Oh man, have we hit 100 likes? Are you kidding me? It's 427. How long have I been doing this?
3 hours and 45 minutes.
TTS text to speech.
Um, any near-term plans to add file tree and editing in line?
Um, depends on what you mean by near-term.
I don't know if you mean like my near-term or your near-term. Um, there are plans to do it. I will say that much. I don't know if they're near-term.
Ryan, talk to me about why you want Oh, text to speech, not speech to text. What do you want text to speech for?
Because I might have something for you.
Um, if you can tell me what you want text to speech for, I can show you something cool.
Maybe that's not what it's called.
talking back in a butler's voice.
Coco the solo maintained Cocoaro TTS fork.
Oh, you want it to read full markdown documents. Um, all right. So what is this a fork of?
Uh it's GPL3 which is why it is uh handled separately.
It's actually very cool.
So there is a um I don't know what they call it a crate or package or something. Um, pretty lightweight honestly that does text to speech pretty well.
Uh, and I have thought about that's why I Oh, there you go. You're using Cocoaro.
There you go. Yeah. So, it's ondevice text to speech and I forked it. um to add a few things and strip out a few things, but it's GPL. Well, it says Apache. I don't think that's right.
Um I think certain parts of it are GPL.
Anyway, I forked it to add and remove some things. Uh and I could ship it at some point. Speech to text would be cool.
Yeah, I don't know.
Um, maybe they changed their license. Is that would be interesting if they changed to Apache. That would make my life a lot easier.
Give me a second, folks. No.
Oh, maybe this is not the one I forked.
How do you tell where this is forked from?
How do you do that in GitHub? How do you tell where it's forked from? Is it like settings?
Yeah, I don't really know.
Hey, we got 103 likes. Hell yeah. And Mark's here.
It's all happening.
Why can I not tell where this freaking thing is forked from?
Oh, maybe I didn't actually fork it.
Yeah, who knows?
Um, but I remember being very careful about uh the license, so I'm pretty sure it's GPL.
That's one thing I've like had to do 104. Thanks, Mark. That's one thing I've had to be really careful about. If I want this to actually be a thing, like I need to have acknowledgements like I ship Jet Brains Mono and Gist Mono and I have to comply by third party acknowledgements like all these licenses cuz if I'm going to like if I'm going to make a go of this, I got to I don't want any I don't want any problems. You know what I mean? Um, so I wish I could figure out what this is a freaking fork from. Am I just an idiot? Or maybe I'm just an idiot. How do you figure out what the hell this is a fork from in the code view? Unless you Yeah, I'm wondering if I didn't actually fork it.
It looks like I did all of this.
Corrected license.
Yeah, I replaced Apache with GPL, which makes me think.
This is infested with something. Here we go. It's in the god dang read me.
All right. Solo term Coco is a solo maintain. Oh, I bet.
GPL.
Why did I think this was Why did I think this was GPL?
Oh, because it is. There's the one I actually forked. It's a GPL license.
Phew. Boy, too much AI, you know, losing my brain.
is the solo maintained forked of Cookar Oaks.
It's very cool.
Anyway, oh that's why it took a thousand years to load. Um I don't want to get into speech to text because I feel like everybody has a speech to text tool, right? Like why would you use my why would you use my speechto text tool that is constrained to solo when you could use super whisper whisper flow monologue and you just use it or as uh that uh person said earlier you could use Jet GPT's speech to text um so I don't I don't really want to get into that game um okay you can push you push pushed. All right, you're dead.
Thanks for all the help.
All right, AMP is doing a final review.
Oh, I found a blocker.
All right, so what did you say about filament over here?
Look, speaking of uh CI, CI is currently running on the Mac Studio right here. So, I'm streaming from the Mac Studio. Um, but the Mac Studio is also set up as a GitHub actions runner because on GitHub the Mac runners are very expensive and I've already spent way too much money on GitHub actions. So, I registered uh I registered my Mac Studio as four separate nodes.
So, it's running four separate uh GitHub actions nodes on here, which is kind of awesome.
And so, what you're seeing here is the uh you're seeing the end to end test.
So, I built a small that's a stress test right there. Um I built a small testing harness that lives inside of Solo itself.
um that gets compiled out uh when it's built for production and it exposes a tiny like HTTP server inside of Solo itself and we can use it to drive end to end tests which is pretty cool.
You don't have the speech input service.
Well, that's a good reason actually.
Mark, what are your what's your filament helpers? Dan was here. Tilly was here.
Got the whole filament crew here.
Your own hosted GitLab. Oh, nice.
Yeah, dude. The freaking with all of these like shards and different configurations. I'm spending so much money on GitHub actions. I think last month I spent like $500 on GitHub actions.
So, I really need to set up more local workers. Okay.
As implemented, we need one new Stripe price. Stripe team price ID.
Solo team is just such a funny It's just such a funny product name that I almost have to go with it. Solo team.
Yeah. Uh, important caveat. Team Checkout does not include the $30 firstear setup fee. I'm actually going to change the pricing cuz it's kind of kind of not doing me any favors either direction cuz you come down here and you're like, "Oh, it's $99.
It's $99 for the first year. I bet it's more expensive after that. This this app sucks." And then you look down here and you're like, "Well, it's $69 after that." Well, Aaron, you should either put that it's $69 and get people to be like, "Oh, it's $69. That's awesome." Or you should just make it $99. Why? What the hell are you thinking? Um, but alas, we don't always make good decisions.
$500 is so much. It's an unbelievable amount of money.
Um, let me just real quick, let's see if we can find this Um, where would that be? Like billing or something?
Where would billing be?
God, even uh look at this.
We're already on our way up.
$20 in GitHub actions just today.
It's brutal, guys. It's brutal.
It's not great.
I gotta move more I got to move more of those workflows onto locally hosted machines.
When solo getit, man soon. I mean having having get inside of solo soon.
Solo git. I don't I don't want to rebuild Git. Um, all right. We need to add this to the filament UI, obviously.
Let's take a look at the filament UI.
Solo term slashadmin.
So, do I have I don't even have stripe prices in here. That feels like a bigger lift.
I don't think I want to do that quite yet for like um for like database school and faster. I have the stripe prices all wired into the admin panel, but I think this one is environment variable driven.
All right, why don't you go ahead and uh flesh out the whole filament admin panel. Save for the Stripe team price ID, we'll just run that through an environment variable. Um but for the license pool resource um and any other affordances we need in filament. Go ahead and wire that up. Oh, I can mark off 100 likes. Hell yeah, brother.
Let's see.
Let's just tell Codeex to do that.
Can you mark the goal of 100 likes as complete? And can you um mark the ad team pricing as active as well?
Let's see. Let's see if Codeex can figure this out.
Wait for it. Wait for it.
Keep waiting for it.
Tada!
We did it.
Thanks, fam.
Shouts out. Shouts out to the fam. Do I sound jinzy?
All right. You claim You claim that the redesign is done.
Now, brace yourself because I don't know if this is going to be any good because we just yoloed it, right?
So, I'm going to use Look at that. Look at that. Look how nice that is. I can just hover over it. It gives me this little link. It says open faster. Because it saw that it's a Laravel project and it picked up the app URL from a thousand years ago.
Hey, we're getting somewhere.
Honestly, honestly, not bad.
Not bad.
Little rough, but hey.
for for for a one shot. It's pretty good.
See, we already have teams. We already have teams built in there.
Honestly, it's not bad.
It's not bad.
I think this That's a little jump scare there with the flash of white.
All I got to say is hell yeah, brother.
Can I get a hell yeah, brother?
It needs a little work. But we we've been knowing this. We knew this, you know.
Not bad, my friend. Um, okay. So, this looks really good. Let us now work through the rest of the pages.
Manifesto legal off series.
Uh no, not Siri. Um yeah, I approve you to keep going.
Let's extract um layouts where necessary and finish up the rest of the pages. And then after that we will um uh what did you feed it to get that? I fed it the design that I paid somebody to make. Um who's Cletus McFarland?
That sounds like an SNL character.
You are still working on canonical ops.
Okay, there are there are a lot of operations that it's got to work through work tree is clean. That means you are dead.
Oh, barnacle boy, do not do not apologize. Um, it is monologue.
That is my one weird trick. Pay somebody talented to do it. So, I paid the same people that did solo, the solo uh site, which is quick 14. I paid them to do the uh faster.dev, the new site.
See, they're really good.
Motorhead. That's funny.
Not cheap.
Not cheap, but worth it.
Yeah, they're good.
All right, let's chat a little bit because I'm about to log off. I'm getting a little I'm getting a little gassed.
Hit me Hit me with your questions before we go. We've been going 4 hours.
What do you got? What do you want to know? What do you want to know?
God, I just love notifications. Oh, can I play you the new I um I made some new notification sounds.
It's probably really quiet.
That's it.
That's it. Those are the sounds.
Oh goodness. Um, so I generated those through 11 Labs and then here are all the like losers.
You like Cloud the best? Yeah, I think I'll I think I'll add these because I think the the default ones that I ship I think those are all um the Mac OS native ones.
kind of like hero.
I'll stick to that one so I don't so I don't trigger you all.
Working on a new green field project at work using a simple API using TypeScript and man rough time that I want when compared to working with a Laravel API. I maintain man TypeScript's a mess.
I don't think TypeScript's a mess. I think the whole ecosystem is a mess. Um, look at that. I freaking love this so much.
Manifesto QA was paused on a Laravel boost approval for inertia route smoke test. I approved it. Yeah, you did. You don't need me. Just do it. Just approve it. What the hell am I here for? Um, that's okay. Sean Washbot, I can't think right now. So, you can't type. I can't think. We're a dynamic duo. I could do different sounds for different alerts.
Um, at that point it gets into the question of like uh configuration.
I don't know. I'm I I'm I'm not there quite yet.
Typescript not having an opinion is an opinion. Hot take. Typescript people love their opinions. Uh, this terminal font is brought to you by Versel. Speaking of Typescript, speaking of a mess. So, I ship a couple of fonts bundled with Solo. These two are free. Jet Brains Mono and Gist Mono. Those are free. I bought for you, dear listener. I bought Mono Lisa and I bought input Mono just for you. So, um, these are these are commercial fonts that have, uh, a lot of commercial fonts don't have redistribution licenses, but Mona Lisa and Input Mono do, so I'm allowed to redistribute those.
And then, yeah, I turn my font weight down. I turn it down, uh, cuz I like it I like it to be nice and lightweight.
All right, you implemented filament.
Nice. Good for you, big dog. All right, what other questions? I'm about to sign off. What else you got? What else you got?
Sweet, sweet, hanging. It's fun to hang out with y'all.
You want to watch some end to end tests run?
Out of interest. What is this? Out of interest do the cost.
Out of interest do the cost. Come on.
Come on. of your multiple AI subscriptions shown your income sales.
Uh no maybe I don't. Yeah. 105. Hell yeah.
Um no. What I showed if you saw the Stripe dashboard earlier that was that's just topline. That doesn't include AI.
That doesn't include designers. That doesn't include GitHub actions. Um, no, that's just that's just topline revenue. Not a question. I'm currently planning a massive refactor of my radio automation systems playout module.
I do love a big refactor.
The Tesla tunnel. Hey, that's cool. Tell Ian he's going to hate that.
What are you using for end to end tests in the app? Unfortunately, a custom harness. Um because uh so I'm using this is a tari application and there's a platform limitation of tari well of it's really webkit doesn't expose like I think it's wdio maybe I don't really remember it doesn't expose like the playright hooks that you need to drive the browser so I had to build a custom harness inside of the app itself itself that gets um compiled out when it's built for prod that um basically gives us a bridge to where the um the command line can call certain commands and do certain assertions within the front-end web view. So that's what you see running here.
Windows baby.
I think the only thing holding me back from Windows right now is fear.
I mean, the the PR is ready to merge.
Um, honestly, I should probably merge it before we get too far. So, I have to rebase the whole thing again. Um, but I think it's I think it's ready. I'm just scared.
What else? What else? What else?
I know. I know. I know. You're right. Of course you're right.
You guys want to see the overhead cam before we go? Tada.
That's so cool.
That's on an iPhone. Uh like an iPhone 12 up there.
All right, nerds.
I'll give you I'll give you three more minutes and then I'm out of here.
Random hot take. Like you want a random hot take from me or you have a random hot take? Yeah. I went to my kids school today and had lunch with them. I took them uh I took them Chick-fil-A.
Orchestrator orchestrating.
Got its timers set up.
You want a random hot take? Um, give me give me give me a give me an area.
Getting a give me a domain. Token Town comes tomorrow.
Ian is always right. That's That's ice cold. That's an ice cold take. Aaron is always right. The playlist is Stream Beats by Harris Heler and this is the Groovy album. This is uh copyright free streamable music domain education. All right, I'll give you a hot take on education.
Um, I think Oh, yeah. You're right. You did say that. Aaron is always right. Ian is always wrong. Yeah, that's true.
Everybody knows that. Uh, education, developer education.
I think the majority of developer education is dead in the way that we used to think of it.
Uh, and the way that we used to think of it, um, the way that we used to think of it was teaching developers how to do implementation.
Um, and I think that's fully dead. I have not abandoned teaching. I don't think teaching is dead. Um, but what I think So, let's take I recently did um it's the wrong browser. Look, this is another computer. How cool is that?
Look at that. That's a MacBook uh that I have back in the closet. Um, so I recently did this one.
Look at that short hair guy. That's a good haircut. I got to do that again. Um I recently did durable objects four hours on durable objects. Um before AI came along I did high performance SQLite my SQL for developers intro to Postgress mastering Postgress.
He was so young and so well rested somehow despite having four kids but there was no AI at the time. So he was well rested.
Um, so what changed? Durable objects is the first post AAI database course that I've done. So what changed in that course? Um, I think the future of education needs to be more not I don't know maybe maybe high level is the right way to do it uh or right way to say it but I think the future of education needs to be more about filling in the user or the students mind about what is possible and less about how to you know structure a react component or whatever. First kid in two weeks. Hell yeah.
That's exciting. You're going to love it. It's the best thing in the world.
Um, so with the durable objects course, it was less about like here's how you um here's how you write the TypeScript code to instantiate a durable object and it's more about here's what's possible with durable objects and let's get our agent to implement this certain thing. But like as you're thinking about the architecture of your app and the primitives and what the platform provides and what you need to do and like you had you kind of have to move up a layer and teach the student this thing exists and when you encounter a problem that is shaped like this you should reach for this solution and how you implement it. You can go down and have your agent start implementing some of that. But if you don't even know that like durable objects are a good solution for this class of problem, you're never going to reach for durable objects. And so I think the future of database school of faster.dev is illuminating the student to when you have a problem that's shaped like this, there is a solution that is shaped like that and less like make sure you put your semicolons here, that sort of thing. Um, so I don't think education is dead. I think it's just a whole lot different.
Um, this is also a good website that I also did not design. There's a hot take for you. All right.
Um, yeah. So Matt PCOX leaned real hard into teaching AI. Um, and I'm doing a little bit of that. That's, you know, we're revamping faster.dev so I can keep the videos flowing over there. Um, I'm torn on that.
just being a pure educator because just being a pure educator I feel like you can lose touch a little bit um with like ground level truth then you end up constructing artififices or like constructing all of these things that are good in theory but like as a day-to-day developer like you may have noticed today I didn't use a single skill and I don't know if that's right or wrong but as a actual like product builder.
That's how I do it. And so now I can take all of this and go back to faster.dev and teach some of those skills over there versus just trying to churn out educational content that is based on something. I don't know. So I like to anchor like when I'm teaching databases, it all comes from lived experience. When I'm teaching AI on faster dev, it hopefully comes from lived experience versus like I think the struggle of a developer educator is are you working on something real or are you just working on like your course platform which is not not to say that that is bad um by any means but I don't know that that always lines up with what working developers um are actually doing.
We got a wild YouTube ad starring Megan Fox at a fake college. I mean, I need to turn off ads, but that sounds better than a hot take on education. Um, so I have a whole series for faster.dev dev planned for it's going to it's going to take me the next two years to complete but it's basically like when you're coding with AI and you see this happen you need to reach for this and so an example is um when you click a button here and something over there changes it might be that you have like a state machine problem like you updated some state here and it got polluted over here or you have like an is error and an is success and an is loading and sometimes they conflict with each other and so I I'm trying to move up into teaching like hey here's what a state machine is and here's when you should tell your agent I feel like we need a state machine here's what a ring buffer is and here's when you should tell your agent I think we need a ring buffer here so I'm trying to move more towards the um when you feel this type of problem, reach for this type of solution because you know if I'm going to reach if I'm going to reach vibe coders or whatever I need to start teaching some more of the fundamental like computer science stuff and I don't have a computer science degree. Um but I know some of these things from my many many years of struggle.
What would Aaron do? Oh, that'd be great. I'd love I'd love a model there.
I would use that. All right, this is it.
Last call. Last call for questions.
Yes. Kind of like those cookbook books.
Exactly.
Here's a problem. Here's a solution.
Totally.
Last call for questions, concerns, or compliments.
Oh, accounting. Left that life left that life far behind me. Anything?
Anything? Anything? Anything?
All right, my friends.
The orchestrators are cooking.
Yeah, the cookbooks are awesome.
Maybe a vibe coding cookbook would be a good idea.
Orchestrators are still cruising. We'll get some solo updates shipped out soon.
We'll get the faster site redesigned shipped out soon.
Thanks for hanging out, Derek. I super appreciate that. Oh, timer just fired.
Somebody's idle, so the orchestrator's going to wake up. You know, we love to see it. All right, my friends.
Going to go cook some pizza. Hell yeah.
Love some pizza. Let's get a slice. All right, I'll uh I'm not going to promise that I'm going to start doing live regular live streams, but I do like live streaming.
But we'll see if we get to a regular cadence. Thanks for hanging out.
We'll see you soon. We'll hit we'll hit you with an overhead.
Goodbye, my friends.
See you on Twitter. X.com the
Videos Relacionados
resume fixed instantly 😭 Comment “app”andI’ll sendyou the link #parakeetaipartnership #resumetips
Ritcareer
686 views•2026-05-31
Re: 🗣️📍theprophedu📍2026 GST 103 CLASS (E-EXAM REVISION)
theprophedu
636 views•2026-06-04
WEB TECHNOLOGIES UNIT-2 | Degree 4th sem BCOM Computers web technologies unit-2 full explanation💯✅
LearnwithSahera
1K views•2026-05-29
More tests are always better? How to use AI to identify tests that bring little value
Alliance4Qualification
335 views•2026-05-29
Search Algorithms Explained in 60 Seconds! 🤖💨
samarthtuliofficial
218 views•2026-06-01
Making Minecraft Clone with C++ & Raylib
PecaCSLive
686 views•2026-06-04
People of Game of Thrones using JavaScript DOM
AltCampus
296 views•2026-05-30
Instagram accounts got PWNed
EricParker
13K views•2026-06-03
