1
The gates are missing buttons
Commit, Accept and Reject are not promises an agent might forget. The commands do not exist on the agent's side at all.
AI can act. Human decide.
The AI does the work. You keep every decision.
A build system where three agents plan, build and check each other — and every point where something could be lost, changed or shipped is a button only a person can press.
- Ideayou drop it
- PlanGrok writes it
- ExamCodex sets it
- You Commitnothing is built first
- BuildClaude builds it
- You Acceptor Reject, with a reason
Six stages. Three agents. Twenty-eight rules. The two blue stops are the ones no agent can press.
The board
Every piece of work walks the same road.
An idea becomes a plan, the plan gets an exam written against it, and only then can it be built. The AI moves the card between stages. Three crossings are yours.
- Commit — nothing is built until you press it.
- Accept — nothing is finished until you press it.
- Reject — sends it back with your reason attached.
The three agents
One job each, and no agent may take another's.
They run side by side in a single session, each watching the board for the work that belongs to it.
- Grok asks about the idea until it is clear, then draws the shapes you choose between.
- Claude builds the one you accepted — and never designs what it is about to build.
- Codex writes the exam before the build, then reads the finished code.
What it remembers
Nothing gets solved twice.
Every fixed bug and every lesson is written down where the agents have to read it before they guess. The rules that govern them live in the same place, read-only.
- Study the log before touching a bug — not after.
- A fix that teaches something general is promoted to a lesson.
- The rules are visible to you, not buried in a prompt.
Four rules that stopped something real.
Twenty-eight of them exist. These four each came from something going wrong once, where writing it down was cheaper than it happening twice.
2
Nobody marks their own homework
A different agent writes the test list, before the build starts. The builder hands it back ticked, item by item.
3
The loop cannot run forever
On the third round of find-and-fix the board stops offering another round and asks you one question instead.
4
The builder cannot check itself
Reviewing is a separate lane. The agent that wrote the code is refused the command that would sign it off.
The rest of the windows, as they actually are.
Overview, the board, Agents Labs and Knowledge are above. Here is the rest — real captures from the running app, not drawings of it.
The machine moved fast. Every decision stayed yours.
That is the whole system. Everything above is the bookkeeping that makes it true on a Tuesday, at round eleven, when nobody is watching.
Back to the top