agent-orchestrator（Agent 协调器）

Agent Orchestrator

Orchestrate complex tasks by decomposing them into subtasks, spawning autonomous sub-agents, and consolidating their work.

Core Workflow

Phase 1: Task Decomposition

Analyze the macro task and break it into independent, parallelizable subtasks:

1. Identify the end goal and success criteria
2. List all major components/deliverables required
3. Determine dependencies between components
4. Group independent work into parallel subtasks
5. Create a dependency graph for sequential work

Decomposition Principles:

• Each subtask should be completable in isolation
• Minimize inter-agent dependencies
• Prefer broader, autonomous tasks over narrow, interdependent ones
• Include clear success criteria for each subtask

Phase 2: Agent Generation

For each subtask, create a sub-agent workspace:

python3 scripts/create_agent.py <agent-name> --workspace <path>

This creates:

<workspace>/<agent-name>/
├── SKILL.md          # Generated skill file for the agent
├── inbox/            # Receives input files and instructions
├── outbox/           # Delivers completed work
├── workspace/        # Agent's working area
└── status.json       # Agent state tracking

Generate SKILL.md dynamically with:

• Agent's specific role and objective
• Tools and capabilities needed
• Input/output specifications
• Success criteria
• Communication protocol

See references/sub-agent-templates.md for pre-built templates.

Phase 3: Agent Dispatch

Initialize each agent by:

1. Writing task instructions to inbox/instructions.md
2. Copying required input files to inbox/
3. Setting status.json to {"state": "pending", "started": null}
4. Spawning the agent using the Task tool:

# Spawn agent with its generated skill
Task(
    description=f"{agent_name}: {brief_description}",
    prompt=f"""
    Read the skill at {agent_path}/SKILL.md and follow its instructions.
    Your workspace is {agent_path}/workspace/
    Read your task from {agent_path}/inbox/instructions.md
    Write all outputs to {agent_path}/outbox/
    Update {agent_path}/status.json when complete.
    """,
    subagent_type="general-purpose"
)

Phase 4: Monitoring (Checkpoint-based)

For fully autonomous agents, minimal monitoring is needed:

# Check agent completion
def check_agent_status(agent_path):
    status = read_json(f"{agent_path}/status.json")
    return status.get("state") == "completed"

Periodically check status.json for each agent. Agents update this file upon completion.

Phase 5: Consolidation

Once all agents complete:

1. Collect outputs from each agent's outbox/
2. Validate deliverables against success criteria
3. Merge/integrate outputs as needed
4. Resolve conflicts if multiple agents touched shared concerns
5. Generate summary of all work completed

# Consolidation pattern
for agent in agents:
    outputs = glob(f"{agent.path}/outbox/*")
    validate_outputs(outputs, agent.success_criteria)
    consolidated_results.extend(outputs)

Phase 6: Dissolution & Summary

After consolidation:

1. Archive agent workspaces (optional)
2. Clean up temporary files
3. Generate final summary:

- What was accomplished per agent - Any issues encountered - Final deliverables location - Time/resource metrics

python3 scripts/dissolve_agents.py --workspace <path> --archive

File-Based Communication Protocol

See references/communication-protocol.md for detailed specs.

Quick Reference:

• inbox/ - Read-only for agent, written by orchestrator
• outbox/ - Write-only for agent, read by orchestrator
• status.json - Agent updates state: pending → running → completed | failed

Example: Research Report Task

Macro Task: "Create a comprehensive market analysis report"

Decomposition:
├── Agent: data-collector
│   └── Gather market data, competitor info, trends
├── Agent: analyst
│   └── Analyze collected data, identify patterns
├── Agent: writer
│   └── Draft report sections from analysis
└── Agent: reviewer
    └── Review, edit, and finalize report

Dependency: data-collector → analyst → writer → reviewer

Sub-Agent Templates

Pre-built templates for common agent types in references/sub-agent-templates.md:

• Research Agent - Web search, data gathering
• Code Agent - Implementation, testing
• Analysis Agent - Data processing, pattern finding
• Writer Agent - Content creation, documentation
• Review Agent - Quality assurance, editing
• Integration Agent - Merging outputs, conflict resolution

Best Practices

1. Start small - Begin with 2-3 agents, scale as patterns emerge
2. Clear boundaries - Each agent owns specific deliverables
3. Explicit handoffs - Use structured files for agent communication
4. Fail gracefully - Agents report failures; orchestrator handles recovery
5. Log everything - Status files track progress for debugging