OpenAI’s Paperclip & Multica imitation?

本站添加: 2026-05-07

[00:00:00]Open AI created something that people are really excited about. It's a way for you to create a ticket, a task in your project management software, and then have agents do the work, and then show you the receipts that they've done the work, presented by Zapier, the AI automation company. As you can see on my screen here, this is the tweet that people have uh have have loved and engaged with.

[00:00:20]This tweet though, I think has a better explanation of what it does. It says you open a task, it automatically spins up an independent isolated Codex agent. It writes its own code, runs its own tests, does its own cross reviews. Damn. Once everything is wrapped up, it hands you a complete evidence package. It is open source, so you can go directly to GitHub, and of course we'll have links to everything here in the show notes.

[00:00:42]Symphony monitors your linear board for work, spawns agents to handle tasks.

[00:00:46]Agents complete the task, provide proof of work. We just talked about it. I've been looking at what people have been saying about it on Twitter and all these different tweets. This person I think is kind of captured it. Been using it since you originally shared it with Linear, and it's been great. Token hungry, but thorough.

[00:01:02]Um Adam, you've installed it, you played with it. This is what it looks like when I'm seeing it on Open AI's screen, but I don't want that. I want your like experience with it. Show us how it works. Tell us about it. Tell us what's good. Tell us what's bad. Do you recommend it? All right. So, this is a pretty cool tool. It's not obvious if you just look at the read me in the repository here what exactly it does.

[00:01:26]So, let me tell you first.

[00:01:28]And actually a good comparison that has a nicer demo is this other tool we've talked about called Multica. Basically, there's all those people out there, and I'm one of them that's chatting constantly with AI. I've got all these chats. I'm talking to Claude. I'm talking to Codex. And in each project, I might have five or six tickets running at a time or or chats here. And it's really difficult to think of these and say, "Well, each one of these probably has its own pipeline of you've got to create the work, you've got to build it, you've got to test it, design it, QA it, and then deploy it." And chat is not a great interface for that. No one would do all of their work in Slack.

[00:02:06]Obviously, most teams are already using a tool like Linear or Jira to organize their tasks. And so, what Multica and now what Open AI has done with Symphony is they've said, "Well, hey, you're already defining these tasks in Linear. So, how about we allow the AI agent to monitor the tasks as they're created, and then to do the work for you?" And it only works in Linear? a Linear set up here. I've got this connected to a project I'm running locally. And I'm going to make a ticket.

[00:02:34]This is just Linear. This is not Symphony. Yeah, hit me.

[00:02:37]>> It does it it only works in Linear? It doesn't work in anything else? Only Linear?

[00:02:41]>> My understanding of Symphony is that it only works in Linear. The way that the install works, it really explicitly is asking for Linear credentials. Although, I can show you the the two ways they have to set it up. We'll we'll go through some of these in in just a minute. And I think especially with option one here, you could connect it to to pretty much any task management system that that you're already using. I used option two, which did seem to be more tied to to Linear. But I but I'll show you those in a minute.

[00:03:08]So, in Linear, I'm just going to kick one of these off so it can run while we're talking.

[00:03:13]Let's see. Let's test.

[00:03:16]Uh create.

[00:03:21]So, normally this would be a ticket.

[00:03:23]Say, I want to add a feature. I want to add the button. I want to fix a bug. You know, whatever ticket. Again, this is, you know, standard software engineering uh software engineering 101. So, but let's to make it like fast because as that tweet you pointed out, it's very token hungry, which means it's also very slow. Uh so, let's to make a faster task. I'm just going to say, you know, create for me a empty markdown file. Or not empty, but just like with one or two sample sentences.

[00:03:51]And I'm going to run that. And what you'll see is when I create the issue, things are going to start happening. So, I want you to I want you to follow along. I'm creating it in the in progress.

[00:04:02]Actually, I'll create it in in to do.

[00:04:04]I'm not going to assign anybody, but it is part of this project. And I'm just going to create the issue.

[00:04:11]So, what should happen >> behind the scenes, and I'm actually going to also show you this. All right, so here's the Symphony dashboard. This is running locally on my machine. Again, I'll show you how this works. But you can see it already picked up one running task.

[00:04:25]It's it's found the ticket, GAT6, which is this ticket right here.

[00:04:30]Inside this ticket, nothing has changed yet. But Symphony's starting to work, and you can already see Okay, here it's streaming a bunch of tokens. I don't really understand why they have the UI like this, but great. We've already burned We haven't done anything. We've burned 37,000 tokens. So, that's not a an amazing start, but it's been running.

[00:04:48]What we'll start to see as it works on it is it's figuring out what is the task, what's the context. It's looking at the code that's in this repository.

[00:04:57]This is a real application.

[00:04:59]Uh it's this application. We use it to manage TV screens. And then inside Linear, as as Symphony keeps working, you'll see Linear will pop into this ticket. It'll move it into the in progress status, and then it'll start to comment, and it'll make a plan, and and all of that.

[00:05:20]It does take a little bit of time. It's already running for a minute and 11 seconds. So far, it hasn't done anything. For comparison, how many tokens should something like this take?

[00:05:29]Well, so so I actually did this as a test. The you saw there was already 300,000 tokens burned here. This was one test that I had run on a very similar ticket. And for that one ticket, it burned 300,000 tokens. I which was just writing a sample markdown file. So, I ran in and just, you know, I want to talk to Codex directly. Here I'm using T3 code to wrap around Codex. And I ask it, "Make a simple markdown file just to test." You know, I wanted to see what would Codex alone do. And that used 14,000 tokens, which is again, we've already like long blown past 14,000 tokens. But here it created the file. In my case, I don't just care about burning tokens. Like this took 10 seconds to write this file, right? In regular Codex. And and now we're sitting over here watching Symphony. Oop, it moved it to in progress. So, you should be able to see Hey, hey, here we go. So, now in Linear, we're we're a minute and a half into we're 2 minutes into into this ticket.

[00:06:30]It's moved the task to in progress. It's Codex created this whole plan. So, this is like a comment on on the on the ticket. And then as it works through the plan, you'll actually sit here and see that it will check off these items individually.

[00:06:45]Which by the way, if this works, I would really like that. I like to see a plan.

[00:06:49]I like to see the plan get checked off as it's doing it. But obviously, you for a simple markdown file, this is way too involved and too slow. Yeah, I mean, I you know, that I think I agree with you, and it's maybe not that fair of a test there. I I wanted especially as I'm just like testing out a tool, I'll often do kind of a hello world example like this.

[00:07:08]Like what's the simplest thing I can do so I can see the tool run all of its motions. And what I would point out is like, "Hey, this task here, it just finished." And now we're up to 582,000 tokens for two markdown files, two like simple markdown files.

[00:07:24]Um it finished. It checked off all the stuff. The the problem is it's it's really slow. Like it's it's hard to do work with a tool that works so slowly.

[00:07:34]Plus, if I'm going to run this on my subscription, dude, I'm I'm going to hit the rate limits so much faster. You know, if right now, I occasionally hit the Codex rate limit, and then it stops me from working for a couple hours using Codex. That's occasional, and I'm only burning 14,000 tokens for a tiny task like this. If you're constantly burning this much this many tokens, I don't know how you're going to be able to get real work done, honestly.

[00:07:59]Okay.

[00:07:59]>> It's done, and you can see it is cool.

[00:08:01]Like I think the the promise here is very interesting, and it moved the task over. This is the the one we just created. It moved the task over, and now it's in the human review column, which is awesome because Symphony Symphony isn't trying to just like deploy this really quickly. It's internal process is I'm supposed to do the work, I'm supposed to test the work, plan the work really well, and then when I'm pretty sure I'm done, move it into a space where where a human can review, and tell the human how to review it. So, that's all that's all being done here, and that's great. Uh but it was a little bit slow, couple minutes, and and burned a lot of tokens.

[00:08:37]But I think there's a lot that's that's actually pretty cool about Symphony, and part of it is how customizable it is.

[00:08:43]Can I show you how some of that works?

[00:08:44]>> Yeah.

[00:08:45]Give me the good, the bad, and the ugly.

[00:08:46]>> Yeah, so installing it is actually a pretty is a pretty cool process. Like often in a repository like this with a tool, you'll see here it's like, "Run this command, then run this command, then run this command." You're like, "I don't know, man. And I don't have that stuff installed." And you end up just pasting all that into an agent anyway. Well, they assume you're going to paste it into an agent. So, they give you two options. It's this, and it's this. For the bottom one, it's basically saying, "We've already built Symphony.

[00:09:15]We know Symphony works.

[00:09:17]You know, here's the interface that Symphony has." And there's a there's a command line interface that they'll give you that that you can run. This is the like the machine that's running, the server that's running on my computer. Um they have all of that already built. And so, instead of building that from scratch, like I want to put that in place. So, I started with this. But the other option to install Symphony is actually to build Symphony from scratch.

[00:09:40]And I'll show you in the file that they reference here, they're really just defining the architecture. There it's not code. There's no code in here.

[00:09:49]They're defining, "Here's how you should build. Here's how you should integrate with with Linear. Here's how you know, these are all like types of things where you would normally be instructing codex on how to build your application. So, they've built that into this markdown file that includes an orchestration layer that says, I know how to move tickets. I know when we would move tickets, how to build a plan. And and in this bottom one, you're just using the pre-built one. And that's what I did. I could show you.

[00:10:18]All I did was I pasted this in. I've got again my existing project that's running and I pasted that in and it started to work. Okay, so the first thing it did was it asked or looked at you know, I'm going to go pull that workflow from Symphony. Cool, it's still working. It's turning through. And then at the end, it actually asked me a couple questions here. I wish I could show you the questions. It doesn't look like I can.

[00:10:43]Nope. But at this point, it starts to ask me questions and and some of those are are actually baked into the the readme that are here. So, this is what if you paste this into your Claude code or codex, this is what codex is seeing. And so, it learns a little bit about Symphony. It learns how to use Symphony. It learns how you should deploy code and push things through linear. And then, instead of you copying and pasting all these commands, Claude just reads this directly and and runs these commands for you.

[00:11:15]And then as you're installing it, it'll start to ask you and it will say questions like, do you want all of these commands? Do you want just some of the commands? Hey, I need you to give me your linear API key. And then it gave a little instruction for here's here's how you would go and get that. And as it asked those questions, as I answered them all, it drops down and says, okay, well, I you know, in the questions, I didn't understand all of them cuz I've never used Symphony before. So, instead of answering, I said, tell me more. Like, tell me about it. And because it's all kind of this agentic flow, it's got the readme, it was able to to describe really well, here's why you would pick this option, here's why you would pick this option, you know, I want to run this. Honestly, I still wasn't sure. I was like, I I just want you to do it.

[00:11:56]Like, give it to me. And that's that's what I told it as I as I gave it the feedback. Like, just do whatever Symphony recommends, whatever OpenAI recommends.

[00:12:04]All right, it cloned down the repository. It built the tool for me. It did ask me to run a couple commands, but then that was it. Like, Symphony was set up. So, I haven't really had to open a terminal. I haven't had to like set anything up. I just answered some questions back and forth and it installed Symphony inside this repository.

[00:12:24]And it did give me a couple things at this point that I had to run in a terminal including providing my linear API key. And it told me how to get my API key. And once I pasted all of that in, that was it, man. I I had I had the terminal running. It's a little the UI here is a little bit broken, but I'm going to reset this key. So, don't even bother trying to trying to steal my linear access. Thanks for I'm glad that you're not making me go and wipe it out. I'll hide these things.

[00:12:53]>> cancel it just as soon as we get off this call. But that's all you do. You you literally just copy and paste those commands and then it started up that server. And the server that was there is is that little interface that I showed you before. And so now, it's a pretty simple system where it says, hey, I'm running. I'm waiting for a to-do item to appear in linear. And when a to-do item appears there, I'm going to tackle it.

[00:13:16]I'm going to go in. I'm going to move it to in progress. I'm going to do the work. I'm going to check out the code, build a plan, and push it through. And that's what you saw run already here.

[00:13:26]Super easy to set up compared to Multica, easier or harder? It was It was Multica's more seems to be more of like a cloud kind of thing. So, in some ways it was more mysterious. Like, they have their onboarding flow, but it wasn't it wasn't very clear or very good. It also wasn't nearly as connected to like, how do I set up the repository? So, I would give Multica maybe a five out of 10 on install and I give Symphony a nine out of 10.

[00:13:53]I would say so, I installed paperclip which I thought it was a real pleasure to install and that would be like an eight, nine out of 10. Very simple to install and it immediately takes you to what feels like a very SaaS-like onboarding process. Okay, so set up all three of these seems to be fairly Multica is a lot more mysterious and gives you less control.

[00:14:15]>> Yeah, Multica is more of a black box and and pretty clearly like they want you to sign up for their platform online. Like, sure, there's an open source version, you can run it locally, but they want you to be buying their their subscription. Like, something I really appreciate about about these open source tools like Symphony is that there's no subscription here. Like, there's nothing. Sure, I'm paying for linear separately. There's Symphony doesn't care. They're just trying to provide for me a tool. And I think the thing they do care about is they want me to burn tokens on on codex, but but I ultimately I really do think they want to be helpful.

[00:14:49]There's just a lot of boilerplate in their system that then runs a bunch of unnecessary work with when you do it.

[00:14:56]Okay, I also I like how it uses linear, uses project management that people are already using. The question that I have is, it's basically its own agent. Where do you give it the rules, the preferences, the way that you work?

[00:15:08]Where is the Where is the tools.md? So, I think that's where you would set up, let's say, you know, Adam, I'm using Jira. I also have like some internal processes here. I've got a team. We do work in a certain way. And then you'd want to you'd want to use this top option because the bottom one is very It's great to get started. It really worked quickly. It was very easy. You could test it out. And I'm excited about what's possible. But you know, one thing I'd point out that doesn't work for me is that great, this thing you saw it grab the task and and move it through.

[00:15:40]If we go and look at linear, one of the things I really don't like, okay, Adam Braken created this issue. That's true.

[00:15:47]You saw me do it. Adam Braken added this comment. No, he didn't. Adam Braken didn't write this. Adam Braken did not move this to in progress. The agent did it. But because it has my API key and because this is the way they set up Symphony by default, it it does it on my behalf. Well, that's not that's not accurate. Like, if I'm working in Multica actually does this well, in Multica, when you when you build a ticket, you can assign a ticket to a person or you can assign a ticket to an agent. So, what I'd actually like to see in Symphony is when I go into to linear and I grab a ticket, I would actually assign to a user that's called Symphony or Adam Symphony or something so that we could be really clear on the back end.

[00:16:33]Oh, yeah, I see what the agent did. I see what the humans are doing and how that conversation goes. Cuz right now, I could comment on this and the agent will see it, but it's going to be a very confusing conversation history cuz it looks like it's just I'm talking to myself. So, if you want to do weird custom stuff like that, which I would argue you should want to do, you would actually need to start instead of building you know, just downloading their existing elixir, which is a programming language, their elixir implementation, you'd actually build your own from scratch. Now, that sounds honestly like it's going to take a lot more time and effort. There's going to be bugs in it. Like, AI never quite gets it right the first time. But then you could ask it, hey, here's all the like custom features I want you to do. Also, I don't want to use linear. I want to use Jira. I want to use this instead.

[00:17:18]Here's how I want it to work. E- even here like something they pointed out is is in the install, it it stops you at some point and says, you have to actually add in extra ticket states like rework and like human review. Those aren't linear defaults, but but the built-in Symphony agent requires these to exist. Well, wait, like my team probably already has, you know, the states that we like. I don't want to invent new states. Or if I do, I want to be in charge. And again, then then I'd be using this top option. But now, instead of a 20-minute install, like you're probably looking at it at an afternoon.

[00:17:55]Okay, so bottom line, it works, not well, slow, burns tokens, good vision of what's possible, needs a lot more work. You'd recommend this, it seems like, for someone who's just looking to experiment and see what's possible, not anyone who's really looking to bank everything on Symphony yet.

[00:18:13]>> Yeah, I I think if you haven't tried a tool like this where you can assign an agent work from cards like from Jira, from Trello, from linear. If you haven't tried one of those, we're not there yet, but you got to try it because for me, this is like it's making me so hungry.

[00:18:30]Oh man, I really want this to work. I now I know how I want it to work because I've tried it and there's some things I like, there's some thing some things I don't like. You know, Andrew, it's back to like remember when touch screens came out? Like, there was a while where everyone didn't really know how to use a touch screen. It didn't know how to use the keyboard versus the physical one.

[00:18:48]And then we all got used to it really quickly. You've got to test this. You've got to feel what it's like to have the agent just show up proactively and do work. And you won't be all the way impressed, but it's going to change how you see the world and it's going to change how you see how AI can plug into your project management, for sure.

[00:19:06]All right, we we're going to keep doing these types of teardowns. Follow for more. I'd love to get more subscribers on this. And of course, we've got the GitHub show that Adam and I do where we do this type of teardown fast for the top GitHub repos of the week. See you in all those and look in the description for links.

[00:19:22]Bye-bye.

[00:19:25]All right.