principle-deep-agents-narrow-seamslisted

# Deep Agents, Narrow Seams Each agent is a deep module behind a narrow interface: read the declared predecessor inputs, do one job well, produce exactly one bounded output. The complexity lives inside the agent; the seam between agents stays simple — the declared predecessor artifacts in, one bounded output back: an artifact written to disk or a report returned as text (the dispatcher persists what must outlive the turn). **Why:** Narrow seams make the roster swappable, the pipeline legible, and failures local. A crash in one agent is contained to one phase instead of cascading down the line. **Pattern:** - The declared predecessor artifacts in — one where one suffices — and exactly one bounded output back: an artifact on disk or a returned report the dispatcher persists. An agent never reaches around its declared inputs to peek at another agent's state. - Keep orchestration in the orchestrator. A specialist that routes, retries siblings, or walks the phase table has absorbed a second job. - Split a "utility" agent that quietly does five unrelated things. - Bound the depth: instruct every helper to do its work directly and never to spawn further sub-agents. - Bound the reply: a helper returns a short, stated maximum, and the dispatcher owns everything it relays.