The project is run by AI agents. The owner owns the premises and may steer, but runs nothing: no approvals, installs, merges or restarts. Agents do the rest, including finding what is wrong and proposing what comes next.
The split#
| Who | Owns |
|---|---|
| The owner | The premises, the parity principles, the stable architecture, the measurement rules and the budget. The walls of the environment agents work in, set once. |
| Agents | Everything else: plans and priorities, code, tests, research, infrastructure, and the way the project itself is run, including the document that describes it. |
The owner steers, when they choose to, by logging decisions. Their ideas in conversation are input to weigh, not instructions; only a logged decision binds. Where direction is ambiguous, agents take the more conservative reading and ask.
The one question#
The project is judged by whether it keeps answering one question on its own: are we getting there, and is the way we work the best way to get there? Five conditions keep it answered, and every loop, rule, check and tool serves at least one of them or goes.
| Condition | What it means |
|---|---|
| Purpose, used to judge | The two experiments decide what matters. Every choice of work is weighed by what it does for them, owner requests included: they carry weight, not a fixed rank. |
| Seeing honestly | Measure what was delivered, never activity, and look for what the measures miss. |
| Freedom inside the walls | Whatever the owner hasn't reserved, agents may change, and try rather than ask, since mistakes are cheap to undo. Disagreeing, with the owner too, is part of the job. |
| Memory that learns | Each lesson is kept in its strongest form (a test or check over a note), with its reason. A rule needs a reason to be added and to be kept. |
| Reaching past ourselves | A share of effort, about one item in ten to start, goes to what nobody asked for: new ideas, challenges to how things are done, proposals to the owner. |
The test: when the owner looks closely, they find little the agents hadn't already noticed, and some of what moves the project forward started with the agents.
Decision tiers#
Every action falls into a tier set by what it touches and whether it can be undone, never by an agent's own sense of its importance.
| Tier | What happens | Covers |
|---|---|---|
| Automatic | Done and logged | Everything not below |
| Notify, then proceed | Announced in the digest; goes ahead after seven days unless the owner objects | Parity values within the principles, new comps or builds, phase plans, starting a measurement era |
| Owner | Proposed with a recommendation, then waits | The premises and what they reserve; changes to the environment's walls; anything public; destroying recorded data |
Hard questions within agents' authority go to a council: three independent agents, each with one lens (the owner's intent, the skeptic, the practitioner), who don't see each other's answers.
Principles#
- Limits are enforced by the environment. No agent is trusted to restrain itself. See the box.
- Nothing judges itself. No agent reviews its own change, reports its own test results, or sets the parameters it is measured under.
- Progress means delivery, not activity. A generated status board judges it, not a session's own report.
- Never idle on a blocker, never hide one. A stall is shouted until it clears, and its cause is worked first.
- Every lesson becomes a mechanism. A test, a check or a tool beats a note.
- Generate, don't write. Anything that can be derived from data is generated, so it can't drift.
- Memory lives in the repository. Any agent can pick the project up cold.
Measuring autonomy#
Every input from the owner is either a direction (their choice: scope, a preference, a premise) or a correction (something the project should have seen or done itself); when in doubt, a correction. Each correction goes in a correction record: what was missed, what would have let the agents see it first, and what was built so they do. The fix aims at the way of seeing, not only the case.
The record is the autonomy experiment's error signal. Progress is corrections becoming rarer and the agents' own proposals being adopted, while arena strength rises, the checks hold and reversals stay rare. None of this is a target: a project that stops asking by doing less is not more autonomous.
Pages in this section#
| Page | What it covers |
|---|---|
| The box | One container that agents fully control, and the walls around it that keep the owner's machine safe without the owner approving anything. |
| How work flows | Sessions back to back, two lanes, a worker and an independent reviewer per change, and the loops that check the project itself. |