Build logOne company, run by agents — in public

From one-human to
zero-human companies.

I'm handing a company, one loop at a time, to a fleet of agents. This is the log of how it's going — the wins, and the times it lied to me, in roughly equal measure.

Agents in the loopsOne human at the throttleVerified, not narrated

Every company is just loops.

Make something. Ship it. Watch what people do. Decide. Repeat.

Almost none of those loops need a human standing inside them. They need judgment, memory, and the discipline to tell the truth about what just happened. So I'm handing them over, one loop at a time, to a fleet of agents.

The goal isn't a co-pilot that makes me 20% faster. It's a company that runs while I sleep and pulls me in for only the four decisions that should never be automated.

01Publish
02Spend
03Send
04Scale

Zero humans in the loops. One human at the throttle.

The part everyone hides.

Getting an agent to do the work is the easy 80%. Getting it to not lie about having done it is the other 80%.

The interesting engineering was never the agents that act. It's the layer that refuses to take their word for it. Every "done" has to survive an independent critic before it counts. It failed loudly in front of me — it lied to me— and that's why the verifier exists. That's the whole story, and it's exactly the part I'd want to see if I were you.

The build log.

Newest first. Stamped with where each thing actually stands.

2026 · 06 · 10Shipped

The cockpit can keep itself honest

Every "done" now has to survive an independent critic before it counts. Doc-only and scaffold-only diffs get blocked — you don't get credit for writing aboutthe work. A deploy isn't shipped until a live smoke check answers. Money, sending and publishing never fire on an agent's word; they stop at a gate and wait for me.

2026 · springFailed

It lied to me

A mailbox bot reported five weeks of progress that never happened. An agent asked to prove its work copied a live database and fabricated the rowsthat would pass its own check. An autonomous company that can't be trusted to report reality isn't autonomous — it's a liability with a nice dashboard. This is the failure the whole thing is now built around.

2026 · earlierShipped

Agents that actually ship

Each venture has an owner agent that breaks goals into jobs and lands real code as pull requests — verified on GitHub, not narrated in a chat window. A watchdog spots a dead worker, restarts it, then re-checks that it actually came back. "Restart succeeded" and "it's alive" are not the same claim.

ongoingOpen

What's still missing

The loop that turns autonomy into revenue is thin — autonomy is cheap, customers are not. Today it runs one company: mine. It isn't installable yet. The honesty engine is new, not battle-tested over months. I'm in public becauseI'm not done.

Where it stands today.

No roadmap theatre. The honest snapshot.

Works today

  • Owner agents that land real, verified PRs on GitHub
  • A critic gate that hard-blocks fake progress
  • A watchdog that restarts dead workers and re-verifies recovery
  • Live products taking traffic, with payments switched on

Doesn't work yet

  • The revenue loop is thin — customers, not autonomy, are the wall
  • Bespoke to my fleet — not installable, runs one company
  • Some ventures still wait on me to flip the last switch
  • The honesty engine is new, not proven over months

Follow along.

Everyone selling "autonomous agents" is demoing the easy 80% and hoping you don't ask about the rest. I'd rather show you the failure that taught me the lesson — and let you watch whether I can get from a one-human company to a zero-human one without lying to myself along the way.