No AI is in charge
The controller is plain code, not an AI. It starts every session, enforces every limit and keeps the record. Each stage gets a fresh orchestrator, and a second model from a different company checks its big decisions.
Every piece of work turns the dial
The builder never grades its own work. A model from another company reviews the plan before anything changes and inspects the real changes after. Disagreement sends it back around, up to a set number of rounds.
- 01Plan Claude
The builder writes a plan. It changes nothing yet. - 02Review Codex
The other model checks the plan against the real code. - 03Lock Bags
The gate approves that exact plan, by its hash. - 04Build Claude
The builder builds and tests what the plan says. - 05Inspect Codex
The other model reads the real diff and tests it. - 06Done Bags
The result goes back to the stage orchestrator. - ↺Revise / Fix
Disagreement sends the plan or the changes back around, up to a set number of rounds.
Roles can swap (Codex builds, Claude reviews), and a loop can run audit-only: one model finds problems, the other confirms them, and nothing in the project changes.
Four stages, four doors
A program runs in stages. Each ends with a report, checked by the auditor first, and a door only you can open. Choose how close you stay: approve every plan, only each stage, or just the last door.
Who answers the knock
The AIs ask before they run anything. Three layers answer in order, so the stream of requests narrows to the few that need a human.
Some commands are blocked outright, and no one can approve them.
Built so it can't fool you
You approve the exact text
An approval names the plan or report you read. Change one word and it no longer counts.
Nothing changes behind your back
The whole project is fingerprinted around every step. An unexpected change stops the program and calls you.
When unsure, it asks
Anything the rules or the judge can't see in full goes to you, never to a guess.
Bounded by design
Limits on loops per program and rounds per loop. Hit one and the program pauses for you.
Never runs on a full head
At 70% of its memory an orchestrator writes a handoff, and a fresh one takes over. Nothing is squeezed.
Your chat list stays yours
Its own chats are labelled and, where the app allows, archive themselves a day after a good finish. Yours are never touched.
A whole AI team. One key.
Brief it once, then step back. Two AIs from two companies build and check each other's work, plain code holds the rules, and you're called in only for the decisions that are yours.