All builds
When an agent hands over a working system, can a person keep it working?
Take over a light installation an agent set up, and keep it working.
Design choice
Make the handover the game: show the actual state, not the agent's summary, and check every repair against futures the player has not seen.
In practice
Judge an agent's work by whether someone else can run it afterwards. Hand over state and checks, not a description.
What I built
- A clock fires every four ticks; delivery lands two ticks later, through a visible queue
- Three seeded faults: a stale queue, a duplicate clock and a cached gate
- Each repair checked against twelve fixed future scenarios, with a counterexample on failure
- A local stdio MCP server and an independent Python checker
Where it stops
Human usability has not been observed; the rehearsal is synthetic and no study has launched. One model recording is shown, which is not a model success-rate estimate.