← Notes

July 2026 · Raian Pollock

The agent constitution

Most companies run agents the way they run interns: one ticket at a time. Write this post. Answer this ticket. Summarize this call. The output is fine and the leverage is small, because the expensive judgment still happens upstream, in a human deciding what the next ticket should be.

That was the right model when agents were weak. It is the wrong model now. As models get smarter, a project brief stops being a help and starts being a ceiling: the agent executes your plan instead of finding a better one.

Think about the best people you have hired. You never managed them by ticket. You gave them a goal, the ground rules, and the context to make calls without you. Then you reviewed outcomes.

Agents are ready for the same contract. It has three parts:

  1. A goal with a number on it. Grow organic traffic to a million visits a month. Cut support resolution time to four hours. A direction and a finish line, without prescribing the route.
  2. A constitution. The operating rules: what the company believes, what it never does, where the money limits are, when to stop and ask a human. Each function gets its own on top of a shared base. The marketing agent's constitution reads differently from the support agent's, the way a CMO's judgment differs from a COO's.
  3. A loop. The agent runs daily, reads its own results, adjusts, and keeps going. You review outcomes and rule-adherence on a cadence, the way you would with any senior hire.

I ran a version of this at my last company. Two people, eighteen countries, eleven languages, and an operating layer of seventy-plus scheduled agents under a written charter, about fifty automated quality gates, continuous evals, and a daily cost ledger. The charter and the gates were the constitution in embryo. The lesson from running it: the rules did more work than the prompts. Prose instructions decay; encoded rules hold.

The hard part is what goes in the constitution. Which failure modes to catch before they burn a week. Where the escalation thresholds sit. What "good" looks like per function, written precisely enough for a machine to execute. That is operating judgment, and it does not come in the box with the agent.

The constitution is the company. Whoever writes the best one runs the cheapest, fastest company in their category.

I write short memos on running companies AI-native. The tooling behind this one is on GitHub, and the practice is at raianpollock.com.