Agent Fleet

Once you have more than one agent, the interesting problems stop being about any single agent: - Something has to own the multi-step flow. Put "implement → review → revise → merge" inside an agent and you have coupled two jobs that change at different rates. Here it is a versioned definition, and it is a **state machine, not a DAG**— the loop back from review to revision is the thing a DAG cannot express. - **Agent work takes minutes to hours.** So nothing holds a connection. Every step is a row with a deadline, a callback token and backoff retries. - **A burst of submissions is not a burst of work.** Each tenant has a concurrency limit; submissions beyond it are accepted and held in `queued`, then started in order as slots free up. Agents that share state corrupt each other under concurrency rather than merely slowing down. - **Discovery has to be scoped.** An agent should find the agents it is allowed to find, and nothing else. Tenancy is enforced in routing, not by a permission check bolted on afterwards. - **A human needs to see where a run is stuck.** Which is a console, not a log file. Agent Fleet is a platform designed to consolidate independent agents into a single, observable fleet, enhancing coordination and efficiency.