Home › Concepts › Multi-Agent Architecture

Multi-Agent Architecture

claude-code's multi-agent system spans from lightweight in-process teammates to cloud-sandboxed remote execution.

Execution Models

Model Isolation Use Case
InProcessTeammate AsyncLocalStorage Same terminal, lightweight
LocalAgentTask Async background Non-blocking sub-agent
RemoteAgentTask Cloud CCR Full cloud sandbox
Worktree Git worktree Parallel code editing
DreamTask Background auto-dream consolidation
MonitorMcpTask Background MCP server monitoring

The Forked Agent Pattern

forked-agent-pattern is the shared primitive. All background operations use the same CacheSafeParams-based fork with cloned mutable state but shared prompt cache prefix. Achieves 92% prefix reuse — spawning 5 agents costs nearly the same as 1.

TeammateTool: 13 Operations

teammate-tool covers the full lifecycle: team setup (spawn, discover, cleanup), membership (join, approve, reject), task coordination (create, claim, complete), communication (send, broadcast, shutdown). Task claiming uses file locking with 30 retry attempts.

KAIROS + ULTRAPLAN: Two-Speed System

kairos handles ambient continuous operations (fast thinking). ultraplan handles episodic deep deliberation in remote Cloud Container Runtime (slow thinking). The community drew the parallel to Kahneman's System 1/System 2.

Coordinator Mode

Encodes orchestration logic in natural language prompts rather than procedural DAGs. Four phases: Research (parallel) → Synthesis (coordinator alone) → Implementation (parallel) → Verification (parallel). Coordinators cannot access filesystem directly — least privilege.

Known Limits