One founder and a fleet of agents
9,274 commits in 90 days, most of them typed by agents. How one person directs a fleet of coding agents, why I publish gates instead of dates, and four decisions that stay with a person.
In this essay5
In the snapshot behind this essay, my Mac held 51 repositories with about 16,000 commits. About 10,300 carried my name, including 9,274 in the 90-day window I counted. These are activity counts, not a measure of product quality or productivity.
The commits on my Mac
I did not type most of those. A commit under my name means I answer for it, not that my hands wrote it. The typing is done by fleets of AI coding agents, Claude Code and Codex sessions, with conductor agents handing out the work. A fleet is many sessions, each with its own task, and I direct them.
What I answer for
The numbers flatter me. 9,274 commits in 90 days sounds like speed. It is also a great deal of material for one person to be accountable for, and accountability does not parallelize. The first thing I open in the morning is the brief in Telegram. It says what happened overnight, and I read it before I approve anything.
I started Mindburn Labs alone. Today we are four cofounders and I am the CEO. “Solo founder” was true when I started. It is not true now. I can describe how I work. How the four of us divide the company is for the others to say.
The order of work
The order I use is the one Elon Musk calls the algorithm: question the requirement, delete, simplify, accelerate, automate last. With agents it matters more. An agent will accelerate anything you point it at, and it will automate a requirement that should never have existed, in an afternoon, with tests. Writing code is cheap now. Deciding that a module should not exist is still the expensive part. So a good share of the fleet is deleting right now: several sessions in parallel, removing what is superseded or duplicated across the repositories and completing only what the next gate needs.
What I own and what I rent
I own the part that makes the company what it is: the decision whether an action is allowed, the receipts it leaves, the commitments, the evaluations. I rent the rest. Identity, payments, hosting and the model SDKs are better made by people who do only that.
Two gaps are still open. I ask every conductor to report progress as a percent of the roadmap. And the fleet runs on my Mac, so when the Mac sleeps, the fleet sleeps.
Gates instead of dates
I do not publish dates. A date is a guess made in public. A gate is a condition that either passes or does not. The gate I am working through now is a coding agent that checks out a task, acts through HELM’s gateway under a mandate a human signed, and leaves a merge with a receipt. Our own development is moving onto it, so the company that builds the firewall also lives behind it.
What stays with a person
Four things stay with a person, whatever the model: hiring and firing, irreversible payments above a limit, legal signatures, and decisions about life and health. Models can prepare any of them. They can draft the contract, and they do not sign it. These are the places where a mistake cannot be rolled back and a person has to be on the record. Inside that line autonomy is earned, per type of action, on a clean record judged by a different model family than the one doing the work.
People call this an AI-native organization. I prefer the mechanism: agents do the work, humans sign the mandates, and every consequential action leaves a receipt. HELM AI OS is the AI operating system I am building so that a company can run this way. I first wrote the idea down as a manual for a company run by agents with one accountable human. It sits in the drawer on the Projects page, and parts of it ended up in HELM.