Transcription
What's up engineers? Indydev Dan here.
As you've heard, the product that went from zero to a billion in 6 months, faster than any product in history, has leaked. Claude Code has been leaked. Every other YouTube tech channel is and will cover the events, the features, the new mythos model, and then they'll give their spicy hot takes. That's not what we do here on this channel. Here we focus on leverage and we scale our compute to scale our impact.
Most engineers are missing the signal of the claw code leak signal you can actually act on. The claw code leak tells me just one thing. And if you're an engineer building with agents, you've picked up on this, too. The agent harness matters a ton. Why is that? It's because the agent harness gives you everything you need to drive agentic results. Deterministic code, token caching, agent orchestration, prompts, skills, and of course, model control. Let's be dead clear here. Without the agent harness, there are no agents, no agentic coding. And that means there is no agentic engineering.
Regardless of how you feel or think about Anthropic, they created the category and they got first movers advantage for it. But let me ask you this. If the cloud code agent harness is worth 2.5 billion ARR now, is it possible for you to take a tool like the PI coding agent and build an agent harness specialized for your domain or specialized for your specific engineering work that captures just fractions of that ARR? Let me spoil this for you. The answer is yes. The Claw Code leak tells us that the agent harness is the product. Of course, they've pioneered a great model that runs on top of it. But as you know, the models are being quickly commoditized, and that means one of the most valuable skills an agentic engineer can learn is harness engineering.
When you stop vibe coding and you start agentic engineering teams of agents and your agent harness, you can solve problem classes, not just one-off tasks. Let me show you what a customized agent harness can really do. We have orchestrators, leads, and workers at the bottom. So, three tiers of agents. I've just modified a few of these models to run Minimax 2.7 and Step 3.5 Flash in their respective teams. We're going to see how they perform right next to Claude Sonnet 4.6 in the domain of building UIs.
So, what are we actually building? What does this agent team surround? So, instead of building a one-off agent to build UIs, what I've done here is designed a system that allows us to build infinite UIs within a consistent brand design. The combination of agents and security is going to be a huge, huge, huge theme over the next few years. I don't know if you've been seeing some of the hacks and the exploits that are occurring. I'll throw a few of those on the screen. The agentic security space is going to be one of the most important business opportunities for engineers, specifically for agentic engineers for the next few years because anyone can sit down right now and just write a prompt to exploit an application. This means that a real black hat hacker can do serious damage.
So what I've done here is I've created a brand around this idea of an agent security command center. Each one of these nodes represents a full user interface. We're inside the brand, we're inside the application branch, and we're inside one of the leaf nodes, one of the versions. So, we've got branded UI. We can scroll down here and just kind of see everything. And the whole idea here is we are looking out for threats with agents in a live real-time way. A lot of the functionality is fully operational. We had our agent teams build this out in a very consistent scaled way that's aligned with our brand. And this is just one of many user interfaces that our system has built. So, we can look at a couple other ones. All about watching out for threats. Here we've actually forked an existing UI design into multiple different versions. So here we have a primary UI type interface, false positives, coverage, performance, deployment control, and then the actual activity logs, so on and so forth.
So notice here how I've upscaled my problem solving. I didn't just sit down in a single cloud code, codeex, open code, insert whatever your agent coding tool is here. I didn't just sit down in a single instance and build out a single UI. You can do a lot more when you scale your work into your agent harness. When you build teams of agents that know how to do the job better than anyone. As mentioned in last week's video, this team is learning as they go. So, these are special types of agents, agent experts. They're keeping track of their own mental models of everything that's gone on. So, when I boot these agents up, as you'll see here in a second, they have already worked on all these UIs. They understand the design. They understand the products. They understand what's going on in their specific domain.
I want to show you a couple big ideas here. First off, the agent harness matters. If you want to push and specialize outside of the core for context model prompt tools, you want to be pushing in to customizing your agent harness. This gives you a massive edge over what everyone else is doing. Spoiler alert, cloud code is everywhere. It is now the mainstream. If you want to push beyond, the next step is to learn how to harness engineer. And the second big idea I want to communicate to you here is scale. One agent is not enough. Spinning up parallel agents with blank memories is an improvement. But what we really want here is teams of agents that remember and teams of agents that can execute like you can. You observe, you act, you learn, and then you iterate.
I have this prompt stored here that's going to create a net new application for us. I'm just going to paste this in here. We'll fire it off and then we'll walk through this process and really understand how powerful this tool can be. So, first things first, you'll notice here that I'm in a chat room with my agents. With our teams of agents, we need to make sure that our input to the system doesn't scale with the number of agents that we have. This directly limits what you can do. So, we have horizontal scaling of all of our agent teams, but we are always just going through a single orchestrator agent. We wrote a prompt. We're doing something really cool here. We'll cover in a moment. I'm running a specialized reusable prompt. So this is a prompt that is truly only going to run in this system that I've built. You can see there there's my chat message, there's the orchestrator, and now the orchestrator is at mentioning the setup team with a bunch of details. It's a full prompt. Let me be super clear here. I have taught my orchestrator how to prompt engineer.
You'll notice some enhancements over last week. We now have a not a to-do list, not a to-do list. We have a till done list. So our agents are going to work until all these tasks are complete. So we have scale on the left and we have structure and communication on the right. And then we have our chat system here in the top. There's an important pattern here as well. Our orchestrator and our leads don't do any work. They're thinkers, their managers, their leaders. Most importantly, they're delegators. Orchestrator delegated to setup and setup lead. Read some key files. It understood what it needed to do. It's already got three trees. And now it delegated its work. So now we're going to start scaffolding. So, the scaffold worker working inside the setup team is starting to get to work. Of course, we're using the new cracked 1 million context window Sonnet and Opus models, specifically the Sonnet model here. But then we have some powerful open source models that we're going to see if they can perform like Sonnet underneath our A and B team here. This is a very very different experience than you're probably used to. I have moved out of the normal distribution of results.
Let me go ahead and just copy this out. This is no cloud code instance. I'm not talking to a single agent prompting back and forth. Our system is starting to build. There's our agent stream and we have these three new UIs. And of course, you can imagine that I have agentically engineered the entire system of agents. We're talking skills. We're talking reusable prompts. We're talking system prompts. We're talking specific tools. I have built this into the agent harness. So, this super team understands how to build UIs inside my infinite UI system extraordinarily well.
Let me be super clear. I'm not all in on the PI coding agent and it's not like I'm fully replacing cloud code. I think that's a mistake. I think in an "and not or" scenario, you want to be using these powerful tools together. 80% of the time I'm spinning up cloud code agents to not work on the actual product or the actual system. I'm using cloud code as a meta builder, a meta agent. When I'm in cloud code, usually I'm building the system that builds the system. If you're a tactical agent coding member, this is going to sound very familiar to you. Because of the fact that you can literally teach your agents to build like you can, there is very little value in prompting your agent to build the thing. What you want to be doing instead is thinking about how you can build a system of agents, specifically a team of agents that knows how to build the application on your behalf, that knows how to operate that domain on your behalf. You can see that happening right here in our system. Setup has completed and now we're deploying our horizontally scaled UI generation team A, B, and C. We can do things in parallel here with these teams. So, UI generation lead A, B, and C. They're kicking off. Of course, they have all been prompted. Again, I've taught my orchestrator how I prompt. It's a huge kind of meta meta theme. My orchestrator knows how to prompt my other agents. And same with the leads. You can see here top to bottom orchestrator prompting team A, team B, team C. And then my leads are picking up the work. And very, very soon, you can imagine what they're going to do. Leads don't work. Orchestrators don't work. They think, they plan, and then they delegate the building to the actual worker levels. We have a three-tier architecture to get maximum value. We have one orchestrator, we have multiple team leads, and then we have as many workers underneath the team that are super hyper specialized to do one thing very, very well. I'm actually running a reduced team version here. I actually have soft validators and hard validators. I have a view generator worker and I have an animation specialist. There's another brand analyst I actually dropped just for the sake of this video so that we don't have to wait here super long for all this work to complete. And you'll notice these X's here. This happens when the agent steps out of its domain. We'll quickly take a look at how I've specialized the system prompt for each agent here in just a moment. But I have built a system that operates this product. The product here in this case is a infinite UI generation system. I can now generate prototypes. I can keep track of the actual working production version of specific systems for many brands. And just to make this super clear, we have a brand here that has several different application pages. Here's the agents app. And if we scroll down here, we have the observability product with three branches. Here we have the dashboard down here. But this goes even higher. If we pull back out here to the workspace level, you can see that we have three brands. We have the Aegis brand for Agentic Security. We have Agentics and Indean. Just playing with a couple brand ideas here. If we click into this at mobile site, we can click in and then we actually have the full versions that are being worked on right now as we speak.
So, this team operates this product. This is like the core of this video. Every week I sit down and I try to communicate as much value as I can to you, the engineer who is operating inside of the age of agents with your boots on the ground. You're working on this stuff every day. You understand that there's opportunity all over. The big opportunity I'm seeing right now is really in the scale of work you can accomplish. Once you take that scale and combine it with trust, building systems that you really trust to do the job, there's a truly absurd level of results you can achieve. If you see that vision, if you understand what I'm saying and you can kind of see how the agent harness plays a critical role in that, make sure you like, make sure you subscribe, make sure you're following the journey. We are going to build with agents until we have agents that work while we sleep. And I'm not talking about simple cron jobs. I'm talking about agents that operate your product end to end. That's been the north star for this channel. It remains the north star. We're not there yet. There's tons of work to be done in building out these specialized systems that operate your domains better than your co-workers, better than you can. That's where this all ends up. So, I hope this all makes sense. I hope you're seeing the vision. Let's go and check on our teams and see how things are going.
We can see our lead team C, which is our Sonnet model doing well, but it looks like we might be having trouble with team B and team A. Let's see what's going on here. Generate timeline stack. No response. No response. Hopefully team Okay, looks like team B might be having the same issue. This could be an issue with the system. It could be an issue with something else. It's not super clear here. Just another reason to have multiple teams and different models running. Looks like Step Flash is having issues. Minimax 2.7. This just adds for that argument of why the agent harness is so important. In fact, I could easily add a model rotation system into this if needed. So, if this model is not working, swap it out. So, there's all types of things you can do when you own the agent harness. I'm really curious to see how our leads and orchestrator puts this all together. The generation on team C. It's very possible that the orchestrator might delegate everything to team C here. We'll see if that actually happens. But yeah, we can see here team B returned nothing. Again, I hope we can see this live on screen that team B and team A is going to report back to the orchestrator. The orchestrator is going to see that team C completed successfully and then they're going to continue working from there. Let's see if this actually comes through.
Okay, interesting here. So, check this out. I need to produce the output to fulfill the task myself. So, the lead actually broke a rule and started writing it itself. Very, very interesting. I wonder if team A did the same thing. The task must complete. Let me update and write the component myself. Okay. So, very very cool. Inside of the system, I have a couple rules that the leads can't or they shouldn't write files and they should always delegate that to the worker, but they realize the workers aren't working. So, I have to do this myself. Very team lead-like. So, it looks like team A and team B did that. I'm curious to see if the validation agents are also going to fail. Let's go ahead and see. Team B, they're going to try to delegate. Okay, no response again from this agent. No response again from this agent. Looks like there's something wrong with the system here. Team C is working there. This is the Sonnet models. Very reliable. We're just going to let our system handle this for us, right? You know, crappy that this happened, especially while filming, but also very valuable to like showcase the importance of having multiple agents that can coordinate, work together, and can solve problems together, right? Actually get to the end state. You can see the UI is getting updated as we speak.
Let me go ahead and showcase what this codebase looks like, right? Let me just highlight a couple things for you. When it comes to the agent harness, we're not going to talk about the exact agent harness details because we covered that last week. What we're going to talk about is an implementation of the multi-team agent harness. I'll make sure to link this video in the description if you're interested. That was a big hitting video that is frankly not getting enough attention, but whatever. I don't control the algorithm. This is the third asset in the Agentic Horizon trilogy. These are member-only assets. First, we covered the CEO agents, lead agents, and now we're talking about UI agents. So, every one of these big ideas we've talked about has been about scaling agents and gaining control over the agent harness. It's almost perfect timing that cloud code source code was released and kind of showcased the importance of the agent harness. You know, once again, we're early to this trend and we're showcasing what we can really do when you own more of the agentic technology. Now, again, that's not to say that cloud code is going anywhere. Clearly this tool is at the very forefront of agentic coding. So I'm holding on to this. I'm doing about 80% of my work in cloud code and then I'm building out systems with it. And then once I have that system, I just use that specialized system that outperforms any single or even multiple cloud codecs Gemini CLI instances. So that's the whole point. That's what I really want to stress to you here.
So let's look at how this team is built. Once you gain control of your agent harness, you can build out any file structure you want. We all know cloud agents, commands, skills, plugins, blah blah blah. When you build your own, we have multi-team here. You saw this last week if you're with the channel. I built whatever I wanted here. So, I have uh the open source agents here for that B and C team. I'll investigate after this video what is going wrong with these open source models. But if we go into our key agents file, you can see all of our agents that we're working with. And if we go into our expertise, you can see there's the mental models of every one of our agents. It's got 7K tokens. The agent operates, controls, and tracks whatever it thinks it needs to. I don't touch this file ever. All I've done here is give this agent a skill which very, very simply tells it how to manage its mental model. But I'm not overprescriptive here. You can see just 75 lines. Our view agent here has just taken full control over what it wants to document, what it wants to track, so on and so forth. This is an agent that learns. It's keeping track of the work it's done. It's keeping track of ideas. If we collapse everything, you can see quite a lot of information. This is yet again another advantage of owning your agent harness.
If we go to the multi-team config min file here, you can actually see the team and how the team is composed. So we have shared context there. Once again, I'm doing this because I own the agent harness. We have paths to some key files, orchestrator path to its system prompt, and then we have our team. So let me just do a quick collapse here. And you can see all the teams in this specific team configuration file. Again, I'm going to be repeating myself here a lot. This idea is so powerful. If you own the agent harness, you can build out any team configuration you want. Here's my full team. And if I collapse out of my full team here, I have quite a few additional agents. So, if you just search for name here, we have 15 unique agents in this. And over here in the min version, which we're running right now in this system here, we have just nine total agents.
Let's go and dial into what a specific agent system prompt looks like. So, let's look at our view generator agent. If we open this up and this is our open-source version because we wanted to change the model. So if we go into this the front matter looks completely different. Once again specialization is at the center of the system. I built out this multi-team agent coding tool to allow for a lot of dynamic customization inside just the front matter. So below we have a normal system prompt with the exception of our injected variables. But you can see here typical format purpose variables instructions. But then the front matter is where a lot of the magic happens. All right. So, first off, expertise file. We are hotloading the skills in to the system prompt, making our agent adhere to them more strongly. You can see there's that mental model skill, teaching the agent how to maintain its mental model, of course, tools. And then another very powerful piece we have here is the domain. So, this is where the agent can actually operate. And for mid to large size code bases, this is very powerful. You can specialize agents to just run on your front end, to just run on your back end, to just run on your database migrations, to just run your DevOps, to just run your billing, so on and so forth. You get the idea. Specialization and that focus context window, one agent, one prompt, one purpose is a massive advantage in these multi-agent team systems.
Let's see what Harvard team's doing here. My leads are taking over. And this is not how the system is supposed to work, by the way. The workers are supposed to get the job done. As in great teams, if you're a lead engineer, you know this very well. If someone, you know, doesn't perform or falls through, it's up to you to get the job done. That's what's happening in the system. Exactly. My A team validation and my B team validation is doing the work themselves. You can see here team lead C is doing some validation work and let's see how they are doing. It responded back to the orchestrator. Two variants need fixes. Let me delegate back to the correct teams. We have a till done system enhanced our agent harness. This system is allowing the orchestrator and the leads to coordinate on what needs to happen. The orchestrator, you can see here pinging team A and team C to fix a couple things. I'm curious if we open up the application what we're going to see right now. Let's refresh. So, it looks like these generations are still coming in. Let's go ahead and let these finish up.
This is what a specialized agent looks like inside of an agent team. What I'm doing here is gaining full control over the core for context model prompt and tools. And to gain maximum control, you really need control of the agent harness. I like the PI coding agent. There are a few others out there that give you the granularity of control that's possible. Now, I have a system where I can come into my multi-team file here and just start building out specialized agent teams. And this is very powerful because it means I can scale it across domains. Fantastic. Again, really centering on that theme of building the system that builds the system. In fact, we're building systems that can be customized to build and run your products and applications and so on and so forth. Another obvious but important emergent property of owning an entire agent team is that you can build prompts against them. When you have a team and you have agents and you can even do this in cloud code as soon as you start building out your own, you know, XYZ.MD MD agents and then the next one you can compose them together. This seems like a very obvious thing. It's a very powerful thing. If you've been building with agents, you know this already. But this has important implications as you start scaling and building agent teams and as you start controlling the agent harness, you can do very very cool, very powerful workflows.
When you start establishing your system for a very simple example, right, here's the generate command. All right, so we're going to generate invariant branches for a tree. You can see all the parameters there. The brand, the product, the tree, the count, and the prompt. And then once again, just consistent prompt engineering. If you've been with the channel, you know this format very well. Variables, instructions, workflow. I'm now talking to teams. And the teams know what they need to do to get the job done. All I have to do is talk to the team. Every prompt, you have to know the communication stream. Who are you talking to? And here, we're always talking to the orchestrator. And so, we're using keywords. Delegate is literally a tool. We're using information dense keywords inside of our workflows to communicate maximum information delegate to X team. Very very powerful stuff here. When we're generating a new set of branches, which is what we ran here, our orchestrator is going to run through this workflow top to bottom. This is a very very important idea. Like 80% of engineering work and of knowledge work, frankly, probably 90% is a workflow. It's a series of tasks. So if you can build teams, if you can build systems of agents that can execute a task top to bottom in a very specialized way that is consistent and will deliver on the results that you're looking for, which is again the point of the agent harness is to increase the trust you have in your system by specializing, you can get some insane results out of this.
All right, so that's this really important idea again that I want to communicate to you here. Let's see what our teams are doing. Okay, we're back into the scaffold team and our orchestrator says we're done. So, let's check out the results. We have three versions and four of our agents completely bombed. The open source models did not run. Maybe that's a configuration problem on my side. I'll figure that out. This is why it's always good to have multiple models in your system and to have teams of agents so that if one agent fails, another agent can pick up the work. Let's go ahead and check out the three new versions of our agent stream prototype UIs inside of our agent security application, Aegis.
Okay, so check this out. Looks pretty good. Even just at a glance here, we're using teleport, which probably uses iframe underneath. Nice glass morphism. Another version, right? Agent stream coming in here. Agents are chatting on the right timeline version. And then, of course, we can see another nice compact dashboard. Let's just click into any one of these. We're just prototyping UIs inside of a product, inside of a brand that is all around building out security systems in the age of agents. I think this is a massive, massive opportunity. If I wasn't heads down hyperfocused on agentic engineering, I would be looking at this as a serious business opportunity. A great place to just provide a lot of value. Business aside, like so many companies are going to need agentic security because the vulnerabilities out there that are going to be exposed by these agents. God forbid someone learns how to build a team of black hat agents with with a system like this. We're talking about nightmarish levels of hacking and hacks and exposed data. This is just a whole field in itself. So, you know, if you've been following the channel and you're looking for a new opportunity, go into agentic security. This is going to be a big field for a long time and a very important field. So, I digress.
So, we have all of our UIs here built out, very specialized, pretty great UIs. And we can, of course, hide our navigation system and actually focus on the raw app. This is a great UI where we have our agents chatting on the right. Our Sentinel agent picked up on an unusual login pattern and that got flagged. And I think systems like this where you have agents always on, always looking at all the logs in your system to look for anomalies and threat detection and perform all of your logs is going to be super super important. You can see here we have some nice filtering and we have great animations. None of these systems work without these powerful models specifically once again looking at claw looking at an entropic looking at the Sonnet looking at these Opus models. What I put together here is that at scale one prompt is not enough. I said that years ago. I was right. Reasoning emerged, thinking emerged. I was earlier to that trend. I was one of the first channels to talk about cloud code and agentic coding. I think I was the first channel to say agentic. So, I'm not flexing here. I'm just saying that I've been focused on the signal for a long time. I've been with this industry since the days of Ader, since the days of GPT3.5. I know where this stuff came from and I know where it's going. The next thing you want to nail is multi-agent orchestration, which increases your total impact that you can have with many agents. One agent is not enough. If you want to push even further beyond a lot of the ideas I talk about on this channel, we're on the edge, guys. We're building things that the big labs will build into their products. Let's make that super clear. Frankly, things that we've done on this channel and maybe some things you've built, too, have been pulled into these big lab products. That's a good thing. It means you're on the right track. Owning the agent harness and having the ability to build your specialized solution for your specialized problem is one of the highest leveraged things any engineer can do right now because when you do this as mentioned you own the core four and then you can scale it up. Big themes to focus on harness engineering, multi-agent orchestration, and then trust and scale. Trust and scale really underlines these first two multi-agent orchestration, harness engineering. That's the system. This is the third and final uh big idea. This codebase is going to be available exclusively for Agentic Horizon members. Let me quickly pitch this. Agentic Horizon is the second part of tactical agentic coding. If you want to tune out now and you've already seen this, you already have the course. Fantastic. This is my take on how you can scale far beyond AI coding and vibe coding with agentic engineering so powerful that your codebase runs itself. These are my handcrafted courses. I built for engineers that ship to production. Okay, we know that AI coding is not enough. Vibe coding is the lowest hanging fruit. And most codebases are now outdated and inefficient. And here's the truth of it all. You are the bottleneck. It's not the models, not the tools, it's not the agent, it's you. It's your abilities. It's your skills. And I say this because I was here too. I realized this and then I pushed into the next frontier. We focus on a lot of key themes here. I'll just point out some of the big ones and then you can decide if this is going to be something that's going to help you or not. Again, we want to build systems that build systems. Solving one-off problems is a waste of your time. Now, there is work that I'll bet you did last week that you could have one prompt, one-shotted into your pipeline of agents, into your AI developer workflow, which is an idea we talk about inside of tactical agentic coding. Coding is dead. Coding by hand is absolutely dead. You want to start templating. You want to template your engineering into your agent system prompts, into skills, into reusable prompts, and then that generates the outcomes you're looking for. So you, your team, and your agents build and improve the templates. It's that meta information, that meta context that then drives the actual building of the application. And you saw that here in the infinite UI. I have a team of agents that operates this application. And then I'm exposing their work through this user interface. I can now prototype any UIs in a brand consistent way with a team of agents to nail it. And despite what you think about the actual UI that is all customizable and can be changed and detailed inside of the system that builds the system. So hope you can like see that really concretely there. And I should probably backpedal a little bit and say that this is not for beginners. If you're a new engineer, if you're a vibe coder, if you don't understand engineering, you know, I really built this for the top 20% of engineers that want to push their abilities to the next level. You saw everything that's going on inside of Cloud Code. If you're watching this leak, you know what's possible. You know the things that they're starting to reach for. You can reach for them, too. But you need to know how to actually control the core for how to actually control your agents, right? Context model prompt tools.
So, what's inside? Let me be super clear about this. There are two courses. If you want to unlock the infinite UI application, if you want to unlock the lead agents codebase, which is what UI agents is based off of, and if you want to understand the kind of first variant of this, the CEO agents, which I'm using now to make strategic high-lever decisions. There's so much compute and intelligence out there in the world, you shouldn't really be making any serious decision without consulting a team of agents. If that sounds stupid, that's fine. This video, my channel probably isn't for you. But if you understand the intelligence that's coming out of these models now, this is probably for you. There are two courses here. Tactical agent coding. Here you make the shift into agent coding. We talk about the eight key tactics of agentic coding. And of course, we build the system that builds the system. And then there's agentic horizon. All right. So this is the upgradable course. If you want these three codebases with next generation multi-agent orchestration and agent experts that learn, you're going to want to pick up both tactical agent coding and agent horizon. I don't want you in the course if you don't want to be there. For that reason, I have a no questions asked 30-day refund before you start lesson 4. Okay? So, this thing is risk-free. If you don't like my style, if you don't like my presentation, that's fine. Bail out. I'll refund you right away. No hard feelings. Okay? I hope you can tell in just the way that I'm talking about this stuff. I've been doing this for a while. In fact, I have proof. Just look back week after week. Every single Monday, I'm trying to share as much information I can. And then every once in a while, I'll put a bunch of next generation ideas into a course like this. This is here for you if you want to gain your edge. If you want to push what you can do with Agentic Coding further for Agentic Horizon members, this third asset, this third codebase is going to be available to you after you purchase. Hit dashboard. After you have both courses, you're going to see this new members asset right here. And if you click this, you'll be taken to the exclusive member-only assets. All right. So, here's a CEO agent. And then you're going to see the multi-team agent coding. And then you're going to have access to the final codebase here, the infinite UI agent coding team, which is a actual implementation of the multi-team agent coding tool.
When I sit down to build with agents now, I think about solving problem classes, not just the individual problem. And the PI coding agent is a big piece of that. And that's because it allows me to control the agent harness. All right, so as mentioned, I'm spending about 80% of my time still working in cloud code. Inside of Cloud Code, I'm building systems that build systems, right? I'm building new teams of agents. I'm building AI developer workflows where we combine agents plus code to outperform either alone. I use cloud code for a lot of in-loop hands-on work and then I'll specialize a system that outperforms it. In a lot of ways, cloud code is becoming my meta agent. The cloud code leak tells us something important. Harness engineering really matters. Controlling what the model does, controlling how you can scale it, controlling the shapes that it takes really matters. I like to use the PI coding agent to specialize my agent harnesses. I like to build agent teams so that I'm throwing more compute at the problem. I'm scaling my compute to scale my impact. No matter what you do though, the theme here is simple. The theme of 2026 is increasing the trust you have in your agents to do larger scales of work over time. If you made it to the end, make sure to like, subscribe, and join the journey as we stack up pattern after pattern, tactic after tactic, and really push what we can do with our agentic engineering. Thanks for watching. Stay focused and keep building.