8.1 Three Multi-Agent Patterns

Claude Code supports three multi-Agent collaboration patterns, suitable for scenarios of varying complexity:

graph LR
    subgraph Pattern1 ["Sub-Agent (AgentTool)"]
        P1[Parent Agent] -->|fork| C1[Sub-Agent]
        C1 -->|return result| P1
    end

    subgraph Pattern2 ["Coordinator"]
        CO[Coordinator<br/>only assigns, never executes] -->|spawn| W1[Worker 1]
        CO -->|spawn| W2[Worker 2]
        CO -->|spawn| W3[Worker 3]
        W1 -->|result| CO
        W2 -->|result| CO
        W3 -->|result| CO
    end

    subgraph Pattern3 ["Swarm Team"]
        T1[Agent A] <-->|mailbox communication| T2[Agent B]
        T2 <-->|mailbox communication| T3[Agent C]
        T1 <-->|mailbox communication| T3
    end

Pattern	Use Case	Communication	Characteristics
Sub-Agent	Single independent subtask	fork-return	Simplest, parent Agent waits for result
Coordinator	Complex multi-step task	spawn + synthesize	Coordinator does not execute, only orchestrates
Swarm Team	Parallel collaborative task	Named mailbox	Peer-to-peer communication between Agents