Keep Agent Work Small Enough to Steer
Agent work is easier to supervise when each job has a visible boundary and the human can intervene before too many decisions accumulate.
Developed from a conversation between Andy David and Pete Winn
Direct messages work well for quick changes, but Pete finds them too low fidelity for important work. Large asynchronous batches create a related problem. An agent can make many decisions while the human is absent, leaving little opportunity to redirect the work as it develops. Pete prefers smaller pieces with regular points for testing, review and correction.
Flight Deck gives each piece a stable place to unfold. One thread maps to one Autopilot session, keeping the conversation and the agent’s work inside the same boundary. A unified inbox shows what has changed, while persistent chats let Pete leave and return without losing where the exchange happened. The work remains visible without requiring every session to stay open.
Documents add another layer of continuity. A document can update during the conversation, then collect comments that Pete and his agent discuss before implementation continues. That resembles a productive human workflow in which a team talks through a shared artefact rather than handing over a large brief and waiting for a finished result. The practical unit is one thread, one session and one developing output that can be reviewed before the next decisions are made.