Essays on AI agents, the systems behind them, and building companies

Write loops, not prompts
A prompt gets one answer while someone is there to ask. A loop runs the same instruction on a schedule, reports every run and waits behind the checks that protect it.

Agent self-reports are not evidence
An agent saying the work is done is a claim, not evidence. Evidence is an artifact that could not exist unless the claim were true.

The twelve-reviewer board
Twelve read-only reviewers in six roles on two model tiers argued against themselves and then against each other, and settled one architecture question in writing.

What evals are for
An eval defines correct outside the thing being measured and checks it after every change, so a prompt edit becomes a measured comparison instead of a guess.

Actor and verifier loops
One thing produces, a different thing checks it against evidence the producer did not create, and the check can say no. That is how unwatched work gets delegated.

Voice-first specs
Intent is dictated in long spoken paragraphs, agents turn it into goals, plans and tasks, and a brake stops them when they start building scope nobody approved.