Architecture

The architecture essays are about the shape of the system around the model: where state lives, who talks to whom, what happens when a run is cancelled or paused or repeated. They are the ones I would hand to someone drawing the boxes before any prompt is written.

A few patterns come up again and again. Shared state between agents wants to be a designed blackboard, not a blob that grew. Side effects want idempotency keys so a retry is safe. Risky actions want an approval gate that pauses the run and resumes it cleanly. Deadlines want to propagate down the whole tree so nothing keeps burning after the user leaves. Subagents want an explicit return contract instead of a lossy summary.

Start with the blackboard pattern, since it is the piece most multi-agent teams are already half-building by accident. Then human-in-the-loop approval gates for the pause-and-resume mechanics, and deadline and cancellation propagation for tearing down in-flight work the right way.