I built a mission control for my AI agents

Ask your AI assistant what you shipped last Tuesday. Watch it guess.

That was my life a month ago. 96 project folders. Five different AI tools. And every single session started the same way: me, re-explaining my own business to a machine that helped me build it the day before. The code was getting written. The context was dying every night. My attention span didn't help either…

So I built a harness. I call it HQ, and it has quietly become the best tool I have ever used. I write about the whole system here on this blog, and the practical field guides live at aifieldcraft.com.

Agents that wake up briefed

Every project in my shop opens with one command. That command builds a cockpit in my terminal (CMUX): the agent in the center, my notes updating live below it, every file change streaming on the left, a rendered preview of whatever I am building on the right.

But the part that changed everything is the boot report. The agent's first move, before I say a word, is to read the operation's files and tell me where things stand. What shipped last. What is stale. What today's card says. Then it asks one question: "What are we doing today?" with three suggestions pulled from my actual plans. My recollection is not needed here.

No re-explaining. No amnesia. The files carry the memory, so any agent, any model, any morning, wakes up already caught up.

A board that runs the company

The whole operation lives on one kanban board with one brutal rule: never more than two things in progress. Every card has a definition of done that a stranger could verify. Every ship gets logged with a lesson, and the lesson is mandatory. Victory laps do not count.

When something annoys me, I log it in five seconds with one command. Later I say "harness retro" and the system turns my complaints into proposed fixes, which I approve or reject. The tool has been repairing itself, on average, within a day of anything fighting me.

A chart room for my own comprehension

Here is an uncomfortable truth about building with AI: you can build things faster than you understand them. This gap in my understanding of what was actually happening kept me up at night.

So the harness now generates what I call the Chart Room. Every project gets explained back to me in plain English at three levels: what it is, how it actually works, and why it exists. The standard is simple. If I cannot explain a project back without looking, the explanation gets rewritten one level deeper. My dashboard is nautical themed, every venture is a vessel sailing from idea to earning…

Loops that grind until the work passes

The newest piece: agent loops. I define a mechanical check, a script that says pass or fail, and point a model at the work. It attempts, the check runs, the failure feeds back, and it attempts again. The model never grades itself. The check decides. When it passes, it commits. When it hits the cap, it stops and tells me the truth.

This matters most for small local models, the ones that run on my laptop with no internet. They get the full training wheels: one small task at a time, skills that encode every mistake their kind has made in my shop, and hard rules they cannot talk their way around. On a train with no wifi, my offline models now do real, shippable work. That still feels like a cheat code.

What it will not do

To be honest, it cannot make a weak model smart, it can only catch the failures before they reach me. It never sends anything on my behalf: every email, every post, every publish waits for my finger. And it did not remove the hard part, which is deciding what deserves to get built at all. That one stays mine.

Why I am telling you this

Because a month ago I did not think this was possible without a team, and it turns out one person with a stubborn streak and a good set of rules can run a whole shop this way. Everything above is ordinary tools: a terminal, git, markdown files, and models anyone can rent or download.

To be honest, I am not even a good programmer, but even though I did not build the tools, I am learning how to use them to do real work in the way my mind works.

I am going to keep writing about the pieces as they prove themselves, including the parts that fail. The deeper write-ups land here on this blog, and if you want the hands-on, step-by-step versions, the free field manuals at aifieldcraft.com are where I put the practical stuff. The next posts get specific.

What would your setup look like if your tools remembered everything and your job was just to decide?