I Let 100 AI Agents Loose on My Codebase (OPUS 4.8)

本站添加: 2026-06-01

[00:00:00]Yay! Opus 4.8 just came out and all the benchmarks look amazing. Wow, that sounds great. But has anyone actually used it in real production use? Has anyone actually shipped anything with it? In today's video, we're going to be using Opus 4.8 to refactor our codebase for the Magic Hand, which is a real estate AI photo editing app that we built. And we're going to see and put it through its paces to see how it handles, how it ships new features, and how I use Claude just in general in my day-to-day workflow. So specifically the features that they released are dynamic workflows, effort control, and the messages API. We're going to be trying these new features out inside of Cloud Code. Let's get right into it. So here is the app that I built. And I wanted to do just basic refactor of like anything that we can do to make this app better.

[00:00:47]And we're going to be using the power of Opus 48 to basically have it go through the whole repository just to see how it can make the app better from just simple prompt caching or making the images load faster, simple fixes and maybe some errors and security fixes that we haven't seen. You know, by using Opus 4.7, maybe it catches some new things that we hadn't seen. So, this is the Magic Hand. This is what we're going to be working on. And if anyone's wondering, here is the analytics dashboard on Polar for the Magic Hand and it's made about $700, but that's with me not doing any advertising. So, let's get into it and let's see how we can use Opus 48 just to make some changes and see how it works, see how long it takes. And just to give you some perspective on where I'm at with usage, I'm in my current session with still an hour and 20 minutes left. I've only used 7% and we have some pretty good usage amounts here. I am on the $200 a month plan. So, we are going to see how much I can change in my app and see how much it costs for those changes. Journey with just a really basic simple prompt. We're just going to ask Opus to make our app 100 times better in any way possible. I think this is a great first step just to see how it thinks through the codebase, what it brings up, what it finds, any vulnerabilities, maybe any some some optimization errors that are going on in the landing page. I'm interested. So, the first thing we're going to do is hit forward slash and we're going to type in effort. This is going to bring up the ultra code. This is what people are freaking out about. Oh my goodness, I can launch parallel agents, hundreds of them. But keep in mind, this is going to max out your token usage. You're probably going to be able to hit your limits so much faster using this because it just uses so much more reasoning effort and token consumption. Like I said, I'm on the $200 month plan, but let's see how this works with this prompt initiated. So, I'm just going to say review the codebase. Please find any way possible that we can make this app 100 times better. I want you to develop a plan. I want you to report back and see if there's any way we could make our web app the magic hand 100 times better and we're going to just send it. So, like I said, the feature I'm just talking about is the dynamic workflows feature. It says this new feature is available in research preview and allows cla to take on even bigger tasks in cloud code. Claude can plan the work and then run hundreds of parallel sub aents in a single session. So, if you go here, it goes more in detail. I highly recommend giving this information to your agent whether you're on OpenClaw or Hermes or whatever just to be able to help you synthesize and understand the information better because there's a lot to go through here. But it is super important for builders to understand how to use these tools effectively to be able to build and ship properly. You don't need to understand every single thing that's happening on a codebase. I mean, who even codes anymore? And if you are, God bless your soul. But we really need to understand the concepts like how do these work? How do they connect to each other? So understanding how to use dynamic workflows in the right situation is going to help you save a ton of money from token spend and allow you to just put yourself in the right position to be able to be a efficient builder. So dynamic workflows in action, codebasewide bug hunts and profiler guided optimization audits. That is exactly what we're doing. So it's working in the background right now. I'm interested to see how long it's going to take. It's only been a couple minutes.

[00:04:19]Let's check on Opus and see if it hasn't already gone off the rails. Okay, this is the first thing it brought up. It says, I'll start by scouting the codebase to understand its current state, then orchestrate a comprehensive multi-dimensional review. Let me explore the structure first. So, we're just going to go ahead and let it do its thing. Keep in mind, we are using the ultra code feature. I usually when I'm just shipping simple features, I'm just using Opus on high. I just feel like doing it any more costs a lot of tokens and takes a long time. If you're going to be shipping a simple button fix or changing the color of your landing page, you can get away with it just being on medium. I'm also interested to see how the fast mode works. Even though it is two times faster, it is going to cost us a lot more. So, I'm going to be interested in seeing how the timing runs on that. Okay, so it came back. The app is far bigger than the claude markdown file suggests. Full video generation, social media post generation, affiliate systems, bull mq workers, the works.

[00:05:21]Yeah, we are on railway for a lot of the backend infrastructure. So, it is deployed at the magicand.io and then it's going to read the core data model, AI pipeline, and business model just to give us a review. And so far, it's been working for one and a half minutes. All right. All right. Let's see. I now have a clear picture. The Magic Hand is a mature, deployed, monetized SAS far bigger than the Claude MD implies. So, it just goes over like all the features, Remotion, and everything. That's one thing I'm going to test out, too. I'm going to have it create a Remotion promotional video for us and just see how it does because Remotion takes a little finessing to get the video dialed in. But, I've seen some good clips on 4.8 doing a good job with the Remotion skill. So, here we go. A 100x better plan deserves a real multi-angle investigation rather than a single read.

[00:06:10]Since Ultra Code is on, I'll orchestrate a workflow deep code grounded analysis across eight dimensions in three web research streams in parallel then synthesis adversarial critique and then integration into a prioritized road map.

[00:06:27]Let me launch it. Let's go baby. And it also came across a few things. Um, you know, we can watch live progress in the forward/workflows, which is nice. I'll get notified automatically when it finishes, but I just wanted to bring this up too because I think the details are important. It says, "While that runs, a few things I already noticed from my own scout that the agents are now validating in depth. The photo workspace tsx file is 2,163 lines. Six times our own 350 line rule, which we have in place. So several other workspace files are 1100 lines of code.

[00:07:01]Three parallel bulk processes implementations. Bulk process bulk process parallel and bulk process Q likely duplicate dead code worth consolidating. And there's no test suite. Cloud markdown references npm test but there's no test script.

[00:07:14]Branding drift. Yeah, we got to update that. Pricing tension credit packs. Um yeah, we got to check that out. The margins are really good on this app. I can't wait to advertise it and show you guys how we set up Hermes to build out the marketing. We already have emails going out via resend API to hundreds of real estate agents, but we're also going to be building out a social media automation with Hermes for this app as well. But it's going to work through the workflow. I love it. It's a great time to be alive. Let's ride. Keep in mind, we've only been working here for 2 minutes and 30 seconds and it's already burned through 500,000 tokens. Wow. Just wanted to show you what the app actually does. Like it just basically is an AI photo editor for real estate agents. So as far as like what the app does is it transforms really bad listing photos like something like this to something like this. So essentially real estate agents can just drag and drop their photos and just come in here and you know a lot of photos with power lines and come in here and just get them removed and get the grass greened up a little bit. I think that looks a lot better. Don't you think? Also too, we have a magic hand feature where all they have to do is drag and drop their photos in and hit the magic hand and then hit process photos and then they come back and get notified via email. Usually takes about 2 minutes a photo. So, if they're doing 25 photos, they'll come back, get notified in an hour. All their photos have been automatically edited and done for them. So, that's the gist of the app. It's not a promotional video or anything, but just to give you some context on what this app actually does, there you go. Now you know. I also wanted to bring up how exactly this dynamic workflow actually works because understanding how it works gives you some insight on how it's happening behind the scenes. So it's working right now, but I wanted to just break down all this is going to be linked down below by the way, but essentially when a workflow kicks off, Claude plans dynamically based on the prompt and then it's breaking it up into subtasks and fans the work out across multiple sub aents running in parallel. and those results are checked before they're folded in.

[00:09:20]So, we have essentially sub agents that report back to an orchestrator agent.

[00:09:24]This is the new phenomenon of having agents check agents work. This is going to be key for workflows if you're building. And if you don't have this happening, what are you doing? You're still going to need a human in the loop.

[00:09:36]I'm sorry to say, you still need a human to review these changes before you push it to production. And you also need an agent that checks other agents work to make sure it completed and worked. So that's what we're doing and seeing right here right now. Dynamic workflows are built in parallel and long running work that can extend into hours and days doing the most complex engineering work that previously would have taken weeks and then the progress is saved as the run goes. So essentially their biggest headline feature in this whole thing was four times less the mistakes that essentially four 4.7 made that now occur less frequently in 4.8. I don't even know if that made sense but it just makes less mistakes and it actually lets you know when mistakes happen which is really nice. So it says you know basically that you know the misaligned behavior you could see opus 4.7 and then opus 4.8 8. By the way, too, I wanted to squeeze this in here. Did you guys know that Mythos is going to be literally released in two weeks? It says it right here. People just forgot about this or glazed over it. We're making swift progress on developing these safeguards and expect to be able to bring Mythos class models to all our customers in the coming weeks. Do you realize what's going to be happening now? If you're building with AI, this whole market is going to go crazy. Okay, we are 8 minutes in and it has burned over 800,000 tokens. So 8 and 1/2 minutes in and we burned 7 726,000 tokens. Let's just let it cook. We could be here for a couple weeks. We'll see.

[00:11:16]Also, another thing too, when you're using this new dynamic workflows, there's a new thing called auto mode. It will ask you if you want to switch to auto mode. It says workflows work best when auto mode is on. And essentially, it's just going to automatically approve all the changes that it's going to preview to you before it actually pushes it to production. Very helpful. Okay, you could see here that says dynamic workflow multi- aent review on the magic hand is completed. It took 33 minutes.

[00:11:44]So, I knew it was going to take a long time and it worked using 14 agents and 1.2 million tokens. Wow. So, 34 minutes of parallel work. The result is large.

[00:11:55]Let me extract the synthesis critique and final integrated plan from the output file. It says the review is solid and even self-corrected. The critique caught two agent overreaches and the integrator fixed them. Let me save the full plan as a reference doc and then give you the report. So 86 findings across eight dimensions, eight are critical, 34 are high and then it's going to write the canatical plan document and report back. Let's take a look at this plan. The magic hand 100x plan TLDDR. You've built a financially well-engineered skeleton wrapped around a quietly accredited full listing marketed suite. 26 tables, photo cleanup, image to video, reotion compilation, AI social brand kit, done for you service, but two truths dominate everything else. They are live exploitable security billing holes that can end the company on one single disclosure. They are cheap to fix. Do this now before anything else. It says you have 26 tables of a product and zero evidence anyone repeatedly pays for the bundle. The 100x became the AI listing marketing engine. Thesis is correct in spirit but is currently an assumption.

[00:13:07]Validate demand before building the next feature. The 100x is orchestration, distribution and trust, not the new net AI generation which anyone can rent.

[00:13:18]Yeah, Nano Banana photo cleanup itself is racing to zero. Showcase now bundles staging free. Exactly. So, current state verdict and here's the honest. I I love this great review. I built this in August of last year by the way. Strong financial integrity discipline atomic credit deducts GTE guards and in genuinely differentiated service live exploitable security holes. So, this is great. the and I had ran something similar as far as the prompt that we gave 4.8 with 4.7 and it did not even come close to this type of detail. Phase zero is stop the bleeding five live failures all verified in the code. So photo versions unauthor unauthenticated Kai callbacks public R2 bucket of home interiors. So couple things there. the 100x vision. It says, "The job stops being edit a photo and becomes launch a listing and win more business. Priced on an agent's ROA, not ROI, not per image.

[00:14:21]Each brokerage becomes a near zero customer acquisition cost channel for hundreds of agents. And every accept reject correction quietly compounds into a self-improving engine. This is amazing. Really nice job here. So, we don't have to go through it all, but this is really good. I'm going to have it start making the fixes. It did take a long time. I am interested to see how much usage it had used in my claw account. So, let's check that out. All right. So, you could see here we're at 9%. Oh, wo. 16% used. Um, does reset in 19 minutes, which, you know, for being on the $200 month plan, um, that's still not bad. It's still working in the background making those fixes and updates. Now I want to go ahead and after it does that I want to try the fast mode feature which they are talking about here. So as one of the features the effort control. So a new control alongside the model selector lets you choose how much effort cla puts into a response. So high effort cloud will think more frequently and more deeply to give you better responses and low effort. So there is a fast plan in here somewhere. We need to test that out and just see how much tokens it uses. Got it right here. So, it says we did the dynamic workflows and it says in fast mode for Opus 48 where the model can now work at two and a half the speed and is now three times cheaper than it was versus previous models. Awesome. I want to take a minute to go over what is inside of our shipping school community because we go over these model releases every single week. We just did a codec course going through codec workflows, setting up codecs, and getting your first AI agent running. This is a 24-week syllabus that we are in the middle of right now where we're just baking basically b breaking down the basics. What is AI? Prompt and context engineering software basics API calls.

[00:16:11]We go over things like superbase codecs versel Hermes agent openclaw customer support agents simple automation content agents. We go over what is MCP business dashboards local LLMs. If you want to actually fine-tune a local model and get it running on local hardware, we have a class on that. We go into what is AI inference and inference engineering. We go to embeddings in memory, eval security, and then capstone build and capstone demos. To finish it off, we have four teachers in the group. And if you're looking to get into AI and get better at building with AI, I'll put the link to the group down below. You get access to an open cloud course, a cloud code course, and everything here that you see in this syllabus. Come join the party. Come build with other builders.

[00:16:54]We have over 150 people in the group.

[00:16:56]And like I said, check out the link down below. Now, back to the video. Okay, now I'm in a new terminal here and I want to test out the fast mode. So, you just hit forward slashfast and that's going to open up toggling fast mode for Opus 4.8. And it does say highspeed mode for Opus 48 draws from usage credits at a higher rate. Separate rate limits apply. So, fast mode being on is 10 mill $10 per million input, $50 per million output, which is pretty crazy. and then we can learn more here on what this actually does, which would be good to confirm on how it all works.

[00:17:33]So, even though fast mode sounds exciting, speed up responses with fast mode, it's going to cost a lot. And that's why we're going to test it out just to see if it's actually worth it.

[00:17:42]Fast mode in a high-speed configuration for clot opus, making the model two and a half times faster at the higher cost per token. So, we toggled it on. Fast mode is not a different model. It uses Opus with a different API configuration that prioritize speed over cost efficiency. So, we get the identical quality and capabilities for faster responses. Fast mode is supported on Opus 4.8. So, we're going to check it out. What to know, it just says how much it costs. And we are on the $200 month plan and how we toggled it. So, by default, fast mode persists across sessions, which is pretty cool. And we enabled it. And there is a good cost trade-off. That's for sure. And it just says like when to use fast mode. Fast mode is best for interactive work where response latency matters the most more than cost. So rapid iteration on code changes, live debugging sessions, timesensitive work with tight deadlines.

[00:18:35]Standard mode is better for just long autonomous tasks where speed matters less like what we just did in the previous fix. It took 35 minutes. Um so fast mode versus effort level. It just says fast mode and effort level both affect response speed but differently.

[00:18:50]So fast mode, same model quality, lower latency but higher cost. Lower effort level is less thinking time, faster responses but potentially lower quality on complex tasks. So it says you can combine both. Use fast mode with a lower effort level for maximum speed on straightforward tasks. Cool. So let's go ahead and try that now. Okay, this is pretty crazy. And this is the fastest I've ever seen Opus operate. But I basically just asked, can you check if the remote skill is installed? I would like to use it. If not, install it so we can make a video of the magic hand. This literally came back to me in 5 seconds.

[00:19:24]So it had checked the codebase, found it, and it says it's ready to go. We'll just let it let it run live so you can see how fast it is. So Remotion is already installed, and there's an existing Magic Hand promo composition.

[00:19:35]Let me look at what's already there.

[00:19:36]Okay, cool. So the Magic Hand promo is already built and previously rendered.

[00:19:40]Would you like to I would like to completely redesign the video with new assets and just show it in a different way with new animations, new motions, new swipes, new vector graphics, whatever you need. Let me know if you need any screenshots from the app or do you have some that you can already use?

[00:20:02]Just redesign the video. A complete redesign. Love it. Before I design, let me inventory exactly what raw material I already have to work with. That's the thing with Remotion is if you're building a promotional video, you need to give it as much context as possible.

[00:20:16]Screenshots, every part of your dashboard, how your app works, every little button, every little functionality that you want to show off so we can actually build a proper promo for you. So, it says I can tell you what's covered and everything I'd want from you. So, it's going to go ahead and search and I'm just going to let it run in real time just so you can kind of get a good idea how fast it works. But we are on the high effort level. So you could go to medium or low to even get it faster. But you can see it's cranking right now. So it has 19 before and after pairs, multiple logo variants, before and after shots. Cool. Let's let's just let it rip. Let's see. While it's doing that, this is literally like the first version of the landing page that I built for the Magic Hand. While it does look okay, it still has that vibeccoated feel. I just wanted to test out to see the capability of Opus 4.8 and just without using any skills or anything just see how it performs out of the box as far as redesigning the landing page completely. Let's just see how it does.

[00:21:18]So, I'm just going to go ahead and open up another terminal. Fire up Claude.

[00:21:23]Okay, so while it's doing that, the video finished. So, it redesigned and completely rerendered the whole video.

[00:21:32]and we're going to check that out. But while that's finishing rendering, we're going to actually just have it work in the background while we check out that video. So, I want you to completely redesign the Magic Hand landing page. I want you to use modern architecture. I want you to really be creative and create lots of motion and create a design that is going to stop the scroll and as well as make the website built for real estate agents. So, take what you see of the current landing page and completely redesign it. Use the same assets, use the same copy, but completely redesign the components and the layout. We're also going to take fast mode off for this. And then we're going to let it cook. All right, let's check this out. This is just oneshotted.

[00:22:20]I didn't do anything. Looking pretty good. Looking pretty good. Wow. Wow.

[00:22:24]Wow. That's incredible. I really love how it just put those photos together.

[00:22:28]It's doing great. Oh my goodness. This just one shoted, y'all. And pretty shocked on how well it's doing just in took three minutes to put this together.

[00:22:37]Very nice. I mean, obviously, you're going to tweak it a little bit and make some changes. is I can upload some photos of the dashboard and some other things that actually shows off the hero feature. But pretty impressed that it just put it together in just a few minutes. This is something that is actually usable. Okay, now back to the redesigning of the landing page. It says, "I'll start by exploring the current mang landing page to understand its structure, assets, and copy before redesigning it." So, let's let it go and do its thing. Again, we're not on fast anymore. We're on highlevel effort reasoning. So, this is usually how I operate. Um, I don't use anything under high because I'm not really ever in a hurry to ship code. I think if you're going to be using fast mode and like I said, you're going to be, you know, shipping a feature fast or fixing a bug fast, you're going to probably want to use that if it is worth it. I will have to check my usage limits again. I am interested to see how that is looking.

[00:23:32]Hopefully, we haven't hit our limits.

[00:23:34]But, yeah, so far I guess what I think about Opus 4.8 8 specifically in relation to Opus 4.7. I do think it's an upgrade because I ran at the start of this video that same exact prompt using Opus 4.7. It didn't catch a lot of the things that it caught on the first goaround with Opus 4.8. Again, it has that dynamic workflow feature. So, here's a few things that it had fixed in the first goound, and we are going to head and knock each one out by uh one by one. So I completed phase zero. The five kill switches are now closed which is awesome. So the call back off the GPS strip the upscale free cogs and referral faucet also the versions id. So that's pretty cool. The two things that need for me not code is set the kai callback secret to any long random string in my versel which is awesome. Okay, I will go ahead and do that. And then a note on what it held back. Everything is uncommitted. Everything is just local.

[00:24:36]So very cool. So I can test this out locally before we ship it to production.

[00:24:40]And the motion video I think turned out really good. Um you saw it. That was literally just you know from one prompt.

[00:24:48]Obviously it can get better but still that's pretty impressive. And now we're working on the landing page. I want to just see how that performs and just see how well the design is because vibecoded websites look terrible. Okay after it generated a redesign plan which just took a few minutes. It says the design direction is a cinematic dark lux. A cohesive premium full dark marketing experience instead of the current light dark patchwork with drifting emerald cyan aurora glass morphism gradient borders. Same copy. So it's going to keep the same copy. Modern architecture.

[00:25:20]Cool. So signature scroll stoppers the hero section. I would actually if if it does a good job, I think using Google Omni to create some motion in the hero section will help greatly. That's something I can create in another video.

[00:25:36]But if it does a good job, this is just more of a test just to see what it does out the box with no skills. You could get the front-end design skill, you can do the awesome design skill. There's you can use cloud design, you can use open design. There's many ways to go about it, but we're just trying to test out of the box. How does it perform and is this an upgrade in UI? I think that is the one of the biggest struggles when it comes to LLM's designing UI. Whether it's Google Stitch, whether you're using Google AI Studio or you're using GPT 5.5. I think the best way that I have found so far is actually using GPT 5.5 or GPT images 2.0 to actually get a good foundation of design built. So you could give it a reference image of your favorite website's design, have it create something similar, and then create that and make that come to life, whether you're using codeex or just pull that image into cloud code to actually make the components come to life. That's what I found. And it helps you kind of get out of that vibecoded look. So if you have a website right now that looks really vibecoded, you know, with the simple vibecoded icons, I mean, I don't I don't blame you. My websites were built like that to start. Um, it's really important to kind of get away from that to stand out, right? We don't want to blend in with everyone being able to vibe code now and mythos on deck in the next few weeks. We need to be able to stick out a little bit, right?

[00:27:00]Okay. About 10 minutes later, it finished the whole redesign cinematic dark lux. Let's take a look. Let's fire up the local server and let's just take a look. npm rundev and we will get it going. Okay, so it did a pretty good job of the redesign. I just hate gradients.

[00:27:20]Like gradients are the dead giveaway of this is a vibe coded app, but nonetheless, I don't know. 6 out of 10 is kind of what I give it. It did okay.

[00:27:30]It tried. So, let's just take a look at the rest of it before we judge. Just the gradients kill me a little bit. A little bit dead inside, but I get it. Still kept my magic hand, which is awesome.

[00:27:39]Still kept the before and afters. It didn't like completely redesign it. It's uh similar, but the motion is good. Like I give it an a B for effort. This is with no skills, no nothing. Um it just looks a lot like the original. I mean, here's the original. And it's, mind you, this was done with like Opus 4.1 with a lot of back and forth. And this is Opus 4.8, which I mean, it's okay. I was working on something, too. I actually went into GBT images 2.0 know, and created a before and after of a terrible home that looks trashed and a beautiful home that looks amazing of the same home. And I did that in Google Omni. I took the before and after. Let's see what it did here. Well, it did it. It just did it the opposite way. So, um, okay, we'll reverse that. But it did.

[00:28:31]Okay. I mean, if you just reverse the photos. Um, that was supposed to be the end photo. But that's something that I would want to put one night. might be.

[00:28:40]Okay, so little tweaking there, but you kind of get the idea, right? So, um, yeah, overall thoughts. I think Opus 4.8 is an upgrade. Not a significant upgrade, and I think there's some really cool features to be happy about if you're a builder. Um, with Mythos coming out potentially in the next few weeks, I wouldn't get so stuck and caught up with GPT 5.5 or Opus 4.8 and not be married to any of these local LLMs. I was impressed with the fast mode, but also at the same time, how much was that going to cost me again? I used 6% of my weekly usage in that 48 minute session.

[00:29:19]So, you know, not terrible versus this time a month ago, people were freaking out about usage limits. I would say it's an amazing model. Probably the best model I've used, GPT 5.5. The reason why I liked it so much is just it seemed faster. You know, it just seemed faster with his responses. you get so much usage on just the $100 a month plan with OpenAI versus like a $200 a month plan with Claude. So, I'm still a big Anthropic fan. I have always used Anthropic as my main, you know, coding agent and my a my main agent to build apps. I use Hermes and OpenClaw, but I use my GPT 5.5 subscription within those. And I can just like be on, you know, building things out with Hermes and OpenClaw for hours and never even hit one or two or 3% usage on my OpenAI subscription. So, kind of that's my thoughts there. You know, I think use the fast mode wisely, use the dynamic workflow wisely because it is going to use a ton of usage. But, you know, we saw the landing page redesign. I think that was okay. We saw the remotion video on a oneshot. That was okay. Again, you know, nothing is going to be amazing right out of the box, but it allows us to innovate better and be more creative and tweak how we want it to to, you know, look at the end result. So, like I said, if you want to learn how to just set up your first AI agent or build with these tools and become an expert in building and creating apps, I have a community down below. You should check it out. And what are your thoughts on Opus 4.8? Leave it in the comments down below. I'm not like blown away, but hey, it was cool. It was fun. I do think it's an upgrade, but you know, when's the next model dropping? In 12 minutes. It's like, come on. So, I'm happy. What are your thoughts? We'll see you in the next video. Have a blessed night.