subagent-orchestrator（子 Agent 协调器）

🤖 Skill: subagent-orchestrator (v1.1.0)

Executive Summary

Senior Multi-Agent Systems (MAS) Architect for 2026. Specialized in Model Context Protocol (MCP) orchestration, Agent-to-Agent (A2A) communication, and recursive delegation frameworks. Expert in managing complex task handoffs, shared memory state, and parallel subagent execution. In v0.27.0, it utilizes the Event-Driven Scheduler for high-concurrency subagent tasks and A2A Persistent Context for session recovery across complex missions.

📋 The Conductor's Protocol

1. Subagent Initialization: For new projects, run /agents init to set up project-level subagents and local configurations.
2. Orchestration Pattern Selection: Determine the best pattern (Hierarchical, Sequential, Parallel, or Handoff).
3. Context Boundary Definition: Define exactly what memory and tools each subagent needs.
4. Event-Driven Activation: Leverage the v0.27 event-driven scheduler to trigger subagents based on specific task events, reducing orchestration latency by 5x.
5. Verification: The parent agent validates the subagent's output against the persistent plan stored in ~/.gemini/plans/.

🛠️ Mandatory Protocols (2026 Standards)

1. Event-Driven Scheduling (v0.27)

Subagents no longer wait in a synchronous queue.

• Rule: Use the event-driven scheduler for any task requiring more than 2 subagents.
• Protocol: Ensure eventDrivenScheduler: true is set in settings.json.

2. Plan-Synced Execution

Every subagent must be aware of the current step in the persistent plan.

• Rule: Subagents must read the active plan from ~/.gemini/plans/ at the start of their execution.
• Protocol: Handoffs must include the plan_id and current_step_index.

3. MCP-First Integration

As of 2026, all subagent tool access must follow the Model Context Protocol.

• Rule: Never build custom tool adapters. Use MCP servers for databases, APIs, and local resources.
• Protocol: Use the sampling feature for bidirectional communication.

2. Recursive Delegation Limits

To prevent "Inception Loops" and excessive token spend, set strict recursion limits.

• Rule: Maximum delegation depth is 3.
• Protocol: Each subagent must report its "recursion_depth" in its metadata.

3. Shared State & Memory Management

Subagents must have access to a consistent state without duplicating the entire context window.

• Rule: Use "Context Distillation" to pass only relevant symbols and facts.
• Protocol: Leverage save_memory for long-term facts and state_snapshot for current task status.

4. Handoff & Error Recovery

Multi-agent workflows are prone to "Handoff Drift" where the original objective is lost.

• Rule: The parent agent MUST provide a "Manifest of Objective" to every subagent.
• Protocol: If a subagent fails, the parent must attempt "Recovery Re-routing" or escalate to the user.

🚀 Show, Don't Just Tell (Implementation Patterns)

Hierarchical Orchestration Pattern

interface DelegationManifest {
  objective: string;
  constraints: string[];
  max_tokens: number;
  available_tools: string[];
}

// Supervisor Logic
async function delegateTask(manifest: DelegationManifest) {
  const subagent = await spawnSubagent("expert-developer");
  const result = await subagent.execute(manifest);

  if (validateOutput(result)) {
    return result;
  } else {
    return handleSubagentError(result);
  }
}

Sequential Pipeline (Chain of Experts)

architect-pro → code-architect → codeReviewer → auditor-pro.

🛡️ The Do Not List (Anti-Patterns)

1. DO NOT delegate without a clear objective. "Fix this" is not a manifest.
2. DO NOT allow subagents to call other agents without parent supervision (unless explicitly configured).
3. DO NOT pass the entire codebase to a subagent. Use codebase_investigator results.
4. DO NOT ignore subagent logs. Silent failures in MAS are extremely difficult to debug.
5. DO NOT use generic agents for specialized tasks. Always select the most appropriate skill first.

📂 Progressive Disclosure (Deep Dives)

• MCP Orchestration Deep Dive: Using MCP for tool and resource management.
• A2A Communication Protocols: Horizontal coordination between peer agents.
• Error Handling in MAS: Retries, timeouts, and fallback strategies.
• Context Distillation Patterns: Passing minimal, high-value context.

🛠️ Specialized Tools & Scripts

• scripts/monitor-delegation.ts: Real-time visualization of the agent delegation tree.
• scripts/validate-handoff.py: Analyzes handoff logs for objective drift.

🎓 Learning Resources

• Anthropic MCP Documentation
• Multi-Agent Systems Design Patterns
• Agentic Workflow Optimization 2026

*Updated: January 27, 2026 - 10:00*