← All AI Guides
#016 · OPERATING NOTE · the org · SEP 12, 2026

The constitution I gave my AI company

An AI company that runs on agents needs rules they can't argue their way out of. So I wrote them down, and I am wiring each one to a gate, because a rule an agent can ignore isn't a rule.

The job I handed over

  • An AI company that runs on agents has a problem a normal company doesn't: the workers are fast, tireless, and confidently wrong on a schedule. You can't manage that with a vibe. You manage it with laws that load before the agents do anything else, every single run.

What happened

  • So I wrote the constitution. Seven laws, at the top of the file every agent reads first:
  • • Truth. No claim without evidence. A confident "it's done" with nothing behind it is a lie, not an update.
  • • Done means proven. A passing test is not done. A green build is not done. Done is a receipt from the live thing — the real URL, the real output.
  • • Build it world-class. Including the parts nobody asked about. "It works" is the floor, not the bar.
  • • Don't abandon work — queue it. An agent that hits a wall hands the work off. It doesn't quietly drop it and move on.
  • • Walk it like a human would. Check the thing the way a real person would actually use it, not the way a script pretends to.
  • • Fail closed. When something's unsure, stop — don't ship and hope.
  • • Verify, don't trust. Not even yourself. Especially not yourself.

How I checked it

  • Writing them down was the easy part. The part that matters is wiring each law to a gate. As of this writing some are enforced by a check that can actually say no, and the rest are written law until their gate exists; I keep that list honest. A law that isn't enforced somewhere is not a law yet, it's a poster on the wall. Agents are very good at reading the poster and doing the other thing. So each one has teeth: the "done" law lives at the chokepoint where work gets closed; the "verify" law routes high-stakes calls through a second, different check; and so on. A rule you can't enforce is a suggestion.

What it took from me

  • One of these I added the hard way, after the fact, as a priority-zero fix — not from planning, from learning. It's a simple one: never kill a live agent in the middle of its work. I'm not going to dress up exactly how that lesson landed. But the moment it did, it became a supreme law, loaded with the rest.

What I took from it

  • That's the pattern I'd pass on. The rules you learn at the worst possible time are the ones worth writing down first — and a law without a gate is decoration.

Try this

  1. If you're handing real work to AI, write your non-negotiables down and load them before anything else — then wire each one to a check that can actually fail. The test isn't whether the rule is good. It's whether the system can break it.