High-Throughput Agentic Engineering by Ex-Meta Principal Engineer

Hi everyone. Or should I say, hi, captains. Many of you might have already watched my previous videos because they got more than a million views. If you haven't yet, my name is Kun. I was an L8 principal engineer previously at Meta, Microsoft, and Atlassian. My previous videos shared a lot of the basics about agentic engineering, so this time I'm hoping to share a more advanced workflow, focusing on high-throughput multitasking through a large number of agents. I work through it by just showing you how I get some real work done. Let's jump in. As always, we will start with a clean terminal. And the first thing I'm going to launch is herder. Herder is basically my new tmux. It's a multiplexer, it manages all my agent sessions and allows me to easily navigate between the sessions. You will see that in a bit in this video. And then I'm going to go into my first mate directory. If you haven't heard of first mate yet, it is an open-source project that I created. It's got quite some popularity. A lot of people are using this. This is basically my workflow. First mate is the essence of my workflow. And the core idea is that I only talk to one agent, which is the first mate, and the first mate will orchestrate all the other agents for me, so that I don't have to juggle between all the sessions and context switch all the time. It's first mate doing that, and you will see that in a bit. The way to set up first mate is super easy. You just clone this repo and launch your agent in this repo. So now that I'm in the first mate directory, I just launch my agent. I use Pi. I use both Pi and Cloud Code. I basically use Cloud Code for whenever I want to use an Anthropic model, like Opus and Fable, and I use Pi for everything else because Pi basically can use any other model, allows me to switch between models very easily as well. I actually, right now, I'm using Grok 4.5. Right now, as of this day, Grok 4.6 is already available, but I actually like 4.5 better because Grok 4.5 is faster and more efficient and more straight to the point. 4.6 is a little bit like Opus 5. It's got this weird personality where it doesn't quite speak like a human. So I like 4.5 better, and that's something I recommend in general, that when a new model comes out, don't assume the new model is always better than the last. Just use it for a bit, and get a sense for how the model behaves by yourself, and look at how other people have summarized their learnings as well. So yeah, this is first mate. And this agent session is where I literally do everything. All my projects, all my tasks, I just use this single agent session to do it. And maybe to show you what that looks like, I'll just ask first mate: "List all the projects that we manage through this fleet." Yeah, and you notice that I use voice input by default, and that's how I work. Voice input is just faster than typing most of the time, like this kind of natural language prompts. The only case where I don't use voice is when I have to type a URL or something like that. Yeah, and you can see first mate has already listed all the projects that I currently manage through first mate. It's a lot of projects, right? This is like, how many? 36 projects in total. So, I do everything across these 36 projects, I do a lot of tasks. Most of these are open-source projects that I published, some of them are private projects as well, and I do everything just through first mate. And to show you how I do that, maybe I just walk through some real tasks. The first thing I probably want to do is this project called backpass. Backpass is a tool that you can use to help you improve your agent.md or cloud.md. The way this tool works is that it will analyze all your agent sessions in the past, and use that to detect what can we improve about your agent.md and cloud.md so that your agents can work better next time. And this tool, I will probably do a dedicated video about this in the future, but right now, I just want to do one improvement, which is something I will just tell first mate: "Hey first mate, previously we planned an idea for backpass, which is to allow backpass to collect agent sessions across multiple remote machines. I think the plan is already there. Can we pull that out, and then dispatch a crewmate to work on that?" All right, so that's how I ask first mate to do one task, and first mate will look into the project and figure out which project this is about and how to find the previously already made plan, and use that to delegate to another implementer agent to do it. You will notice that first mate agent is already working, but we don't see any tool calls, we don't see anything in this Pi agent terminal, right? And the only thing we see is this little boat that is floating around, and sometimes the agent chain of thoughts is showing up as well, but it quickly gets dismissed. This is done by a custom Pi extension that I built. It comes with the first mate repo, and the reason I do this is that I really don't need to look at what the agent is doing. It's making all those tool calls, it's reading files, it's doing this and that. I don't actually need to know that. Most of the time, those things are distractions and noise. So I built this little custom extension called calm. This calm command in Pi will basically allow me to hide all those noise and just show this little boat that allows me to know the agent is working, it's not stuck. If I do want to see what kind of tool calls were being made, I can just type slash calm to toggle these details. So now I can see, "Oh, it's making this tool call, it's spawning a crewmate." So yeah, that's what the agent is doing. But most of the time, I just turn on calm mode and I don't need to look at all those noise. And then that allows me to focus, that allows me to actually think about what do I actually need to do next, right? So that's what's important, not looking at the tool calls and everything. So one thing I do want to show here, okay, first mate said it's found the plan and dispatched it. So now we can see there is a crewmate here, right? In the left-hand side panel where we can look at all the sessions, all the agents. We can see here there's a backpass SSH. This is basically another terminal tab where another agent has spawned up and is running Opus 5. It's got the requirements, and it's now building the thing. It's building what I asked, it's building what I asked. This is basically a crewmate that spawned up by first mate earlier. And the reason we are using Opus 5 here is that if we look, if we toggle calm mode off, and let's see if we can find the previous tool calls. We can probably see that before first mate decided to launch the crewmate, it called a few things. It called this Quota Axi and it used a skill called Quota Array Dispatch. This is something that's really important and that's why I want to talk through this. First mate, when first mate is about to dispatch a crewmate to do something, it will look at a bunch of rules to decide which agent and which LLM to use for that crewmate. So, yeah. And I actually don't write this rule file by hand. I ask first mate to do this. I just tell first mate what my preferences are and first mate will configure this file for me. And the rules, we can look through this. For new feature development work on this iOS app that I have, always use Fable. Fable is like really good at building good UI, and this iOS app actually is a paid app, I really want to make sure it's taken good care of, so I use Fable for that. And the task requires generating images. If it needs images, then I use Codex as the harness and then GPT-5.6 Soul. Codex has a native tool for generating images because OpenAI has the image generation model, right? So I use Codex to do that. There's also when the task is technical product design, architecture, or planning work that is genuinely very difficult. So it's a difficult planning work, I use Fable, I use Kiwi K3, I use Astra. So I basically allow these three models to do this kind of very complex and ambiguous planning for me. And there's also "why" as well, so it helps the model make the decisions. And then if the task is a simple bug fix whose root cause and expected behavior are already well-defined, then I use Luna, I use Sonnet, I use Cursor, Grok 4.6. And then the default, if a task doesn't fall under any of those rules, then the default is Cloud Opus and this Cursor and Grok 4.6 High. So that's basically my rules for how I dispatch the crewmates. And when you see, I have multiple models defined here, right? So when this rule matches, there is three models that's available, and when this happens, the first mate will look at which model has the most quota available, and choose the one that has the most quota. So I'll show you how I look at my quota as well. I use this tool called Quota Axi, that's another project I built. It can give us this data where I can just easily monitor how my quota is doing. I have a subscription from Cloud, I have a Codex, I have a Cursor, I have a Grok as well, and this is where they sit currently. I actually took a vacation right before this and I have a ton of quota available, which is awesome, so I can build a lot of stuff. And this Quota Axi is actually, this TUI is for human, but the default CLI is for agents to use. So first mate will call Quota Axi to get access to all my quota information and use the model that has the most quota available, so that I don't waste any quota. So this is like really useful because otherwise I have to constantly look at this myself, and with people like Tible constantly resetting Codex quota, it's really hard to keep track of that, right? So that's why I use this way, let first mate to manage all this dispatching for me. And right now, first mate is already working, but we don't see any tool calls, we don't see anything in this Pi agent terminal, right? And the only thing we see is this little boat that is floating around. And sometimes the agent chain of thoughts is showing up as well, but it quickly gets dismissed. This is done by a custom Pi extension that I built. It comes with the first mate repo. And the reason I do this is that I really don't need to look at what the agent is doing. It's making all those tool calls, it's reading files, it's doing this and that. I don't actually need to know that. Most of the time, those things are distractions and noise. So I built this little custom extension called calm. This calm command in Pi will basically allow me to hide all those noise and just show this little boat that allows me to know the agent is working, it's not stuck. If I do want to see what kind of tool calls were being made, I can just type slash calm to toggle these details. So now I can see, "Oh, it's making this tool call, it's spawning a crewmate." So yeah, that's what the agent is doing. But most of the time, I just turn on calm mode and I don't need to look at all those noise. And then that allows me to focus, that allows me to actually think about what do I actually need to do next, right? So that's what's important, not looking at the tool calls and everything. So one thing I do want to show here, okay, first mate said it's found the plan and dispatched it. So now we can see there is a crewmate here, right? In the left-hand side panel where we can look at all the sessions, all the agents. We can see here there's a backpass SSH. This is basically another terminal tab where another agent has spawned up and is running Opus 5. It's got the requirements, and it's now building the thing. It's building what I asked, it's building what I asked. This is basically a crewmate that spawned up by first mate earlier. And the reason we are using Opus 5 here is that if we look, if we toggle calm mode off, and let's see if we can find the previous tool calls. We can probably see that before first mate decided to launch the crewmate, it called a few things. It called this Quota Axi and it used a skill called Quota Array Dispatch. This is something that's really important and that's why I want to talk through this. First mate, when first mate is about to dispatch a crewmate to do something, it will look at a bunch of rules to decide which agent and which LLM to use for that crewmate. So, yeah. And I actually don't write this rule file by hand. I ask first mate to do this. I just tell first mate what my preferences are and first mate will configure this file for me. And the rules, we can look through this. For new feature development work on this iOS app that I have, always use Fable. Fable is like really good at building good UI, and this iOS app actually is a paid app, I really want to make sure it's taken good care of, so I use Fable for that. And the task requires generating images. If it needs images, then I use Codex as the harness and then GPT-5.6 Soul. Codex has a native tool for generating images because OpenAI has the image generation model, right? So I use Codex to do that. There's also when the task is technical product design, architecture, or planning work that is genuinely very difficult. So it's a difficult planning work, I use Fable, I use Kiwi K3, I use Astra. So I basically allow these three models to do this kind of very complex and ambiguous planning for me. And there's also "why" as well, so it helps the model make the decisions. And then if the task is a simple bug fix whose root cause and expected behavior are already well-defined, then I use Luna, I use Sonnet, I use Cursor, Grok 4.6. And then the default, if a task doesn't fall under any of those rules, then the default is Cloud Opus and this Cursor and Grok 4.6 High. So that's basically my rules for how I dispatch the crewmates. And when you see, I have multiple models defined here, right? So when this rule matches, there is three models that's available, and when this happens, the first mate will look at which model has the most quota available, and choose the one that has the most quota. So I'll show you how I look at my quota as well. I use this tool called Quota Axi, that's another project I built. It can give us this data where I can just easily monitor how my quota is doing. I have a subscription from Cloud, I have a Codex, I have a Cursor, I have a Grok as well, and this is where they sit currently. I actually took a vacation right before this and I have a ton of quota available, which is awesome, so I can build a lot of stuff. And this Quota Axi is actually, this TUI is for human, but the default CLI is for agents to use. So first mate will call Quota Axi to get access to all my quota information and use the model that has the most quota available, so that I don't waste any quota. So this is like really useful because otherwise I have to constantly look at this myself, and with people like Tible constantly resetting Codex quota, it's really hard to keep track of that, right? So that's why I use this way, let first mate to manage all this dispatching for me.