Back to case study
Hub-Spoke replay

OpenClaw Multi-Agent Orchestration

Multi-agent reduced to three MCP primitives (spawn/send/history), with six modes on top (Hub-Spoke/Pipeline/Hierarchical/Routing/P2P/Fleet). Understanding Hub one-directional dispatch, the subagent-layer sessions_send ban, and why P2P has zero production cases.

A Hub-Spoke code review: the lead spawns three review subagents, an upward sessions_send is shown bouncing off the subagent boundary (blocked), and results flow back only via sessions_history.

OpenClawMulti-AgentMCPHub-SpokeOrchestration
OpenClaw Multi-Agent Orchestration

Why this local version exists

The three MCP primitives (spawn/send/history), the six modes' token multipliers, and the subagent-layer sessions_send ban (Issue #23359, the root of P2P's zero cases) come from the "OpenClaw 多Agent系统入门" course. No live OpenClaw runs here.

Interactive Preview

Hub-Spoke multi-agent code review

Replays OpenClaw multi-agent: a lead spawns 3 review subagents; results flow back only via sessions_history — a subagent's sessions_send is banned at the tool-registry level.

Hub-Spoke topology (3–15× tokens)

leadsecurityperformancereadability

Hub single-directional dispatch: lead delegates and collects; subagents can't talk directly (P2P has zero production cases).

1. lead sessions_spawn three review subagents (parallel)

2. subagents work in isolated contexts

3. a subagent tries sessions_send to lead → blocked (tool not registered)

4. lead collects via sessions_history (or reads shared/signals)

5. lead synthesizes the three reviews

Three MCP primitives (DB analogy)

sessions_spawnINSERTcreate a sub-agent session
sessions_sendUPDATEinter-node msg (lead-only; banned at subagent layer)
sessions_historySELECTcollect sub-agent results

What to try

Run the orchestration and watch the lead spawn three subagents in parallel.

Watch the upward sessions_send arrow turn red and bounce — banned at the subagent layer.

See results flow back only via sessions_history before the lead synthesizes.

What this demo proves

You design multi-agent systems by topology and token cost (3–15×), not by piling on agents.

You understand the Hub one-directional architecture and why P2P has zero production cases.

You know the real engineering pitfalls (misplaced allowAgents, rate-limited polling, lost completion).

Primitives

sessions_spawn (INSERT) · sessions_send (UPDATE, lead-only) · sessions_history (SELECT)

Modes

Hub-Spoke / Pipeline / Hierarchical / Routing / P2P / Fleet (1.5–20× tokens)

Best signal

Evidence-driven orchestration design — knows what doesn't work too