A studio run by AI agents, with a gate for every change
Our engineering is done by AI agent sessions working in their own workspaces, from a written plan, against checks they cannot change. People decide what to build and approve what ships.
Propose, approve, build, prove, close
- 1
Propose
An agent writes the change as a design and a plan: what, why, which files, which tests. Nothing is built yet.
- 2
Approve
A person reads the plan and rules on it. The rulings are recorded with the plan, so the next session can act on them without asking again.
- 3
Build
A fresh session implements the plan task by task in its own workspace, writing the failing test first. Another session reviews each task before the next begins.
- 4
Prove
The gate runs. Its result is the only evidence that counts, and an agent cannot edit the rules it is judged by.
- 5
Close
The change lands, the record is kept, and the ledger says what was decided and why.
What keeps it safe
The agent cannot change its judge
The rules an agent is checked against, and the level of autonomy each path allows, are propose-only for the agent. A person changes them.
Levels of autonomy
Every path in the estate has a level: what an agent may do alone, what it may propose, and what it may not touch.
Plans a stranger can run
A plan is written for a session with no memory of the conversation: exact steps, exact code, and a ledger that outlives the session.
Two seats per task
One session implements, a fresh one reviews, before the next task starts.
Where people are
People write the brief, rule on the plan, approve copy and content, and decide what goes live. Agents do the rest and show their evidence.
Read how we work
Refuse, don't warn
A build that warns about a leak will ship the leak. Make every check a refusal with a name.
Learn more → about Refuse, don't warnA plan the next session can run
A plan for an agent with no memory of the design conversation: exact code, a brief per task, a ledger.
Learn more → about A plan the next session can runTwo seats per task
One session implements and a fresh one reviews, before the next task starts.
Learn more → about Two seats per task