Agent Teams — Multi-session coordination

🌐 日本語

NOTE

When one session — even with /compact and /clear — cannot hold a multi-week task without quality decay, the answer is not "a better session" but several persistent sessions that talk to each other. Where Part 5's Subagents are delegated children, Agent Teams are peers with their own lifespans.

How it works

• Multiple persistent sessions, each with a bounded role (backend / frontend / QA / docs, or by stage / layer / feature).
• Sessions exchange messages via direct messages, a shared task queue, or an artifact store — chosen by how synchronous the work is.
• An optional orchestrator role routes work and arbitrates conflicts; it is a peer with a specific responsibility, not a structural parent.
• Durable state lives in the artifact store, not in any single session's history — so any session can be /clear-ed and rehydrate from artifacts.

Related chapters

• Part 10 — Multi-Session Coordination (overview)
• Subagent vs Agent Team
• Session Boundary Design
• Peer Messaging
• Long-Running Tasks

Structural problems it addresses

Agent Teams hit three problems at once — the underlying remedy (keep every session small) is exactly what each one individually demands.

• Context Rot — root-cause fix. No single session ever holds the whole project.
• Lost in the Middle — each session's history stays short, so the U-shaped recall curve never forms.
• Priority Saturation — responsibilities split across sessions; each carries instructions for one role, not all of them.

See also

• Topic: Skills & Agents — start here if a single-shot Subagent is enough; only reach for Agent Teams when it isn't.
• Topic: Plugins — team-shared CLAUDE.md, Skills, and Hooks can be packaged as a plugin and applied to every session in the team.
• Feature Index