Requesting Status and Trajectory Persistence

Two ways to learn about the progress of a child agent. Via messages: ask “how is it going?” asynchronously or receive proactive status_updates in the bus; express the state as a single dictionary (executing/input required/completed/failed - as in A2A). Via file: the child agent serializes the trajectory to JSONL in real time (analog of pre-record log - state is always reproducible, crash recovery, audit, transfer to another agent) or maintains a lightweight progress.md. Hang detection: the last record time does not change for more than N minutes - timeout safeguard. The trajectory is the entire state of the agent: the prefix is defined by the code, there is no runtime state besides the trajectory.

Related: [Inter-agent communication], [Multi-agent interaction without shared context]