# Principle Deep Agents Narrow Seams

> Keeps agent interfaces narrow and internal work deep. Apply when defining an agent, subagent, or dispatch boundary.

- Skill: `bostonaholic/principle-deep-agents-narrow-seams` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add bostonaholic/principle-deep-agents-narrow-seams`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bostonaholic/principle-deep-agents-narrow-seams/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: bostonaholic (https://skillmd.com/u/bostonaholic)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bostonaholic/principle-deep-agents-narrow-seams

---


# Deep Agents, Narrow Seams

Give each agent one job behind this interface: the declared predecessor artifacts in, one bounded output back—an artifact written to disk or a report returned as text.

- Use one predecessor where sufficient; never inspect undeclared agent state.
- Let the dispatcher persist returned reports that must outlive the turn.
- Keep routing, sibling retries, and the phase table in the orchestrator.
- Split utility agents that perform unrelated jobs.
- Require helpers to work directly and never spawn further sub-agents.
- State a short maximum for helper replies; the dispatcher owns relayed content.