orchestrate-agents（协调 Agent）

Orchestrate Agents

Complex workflows across multiple agents.

Orchestration Patterns

Pipeline (Sequential)

Agent 1 → Agent 2 → Agent 3

Each depends on previous output.
Use for: Analysis → Design → Implementation

Parallel + Merge

[Agent 1]
[Agent 2] → Merge → Final Agent
[Agent 3]

Independent work, combined results.
Use for: Multi-area search → Synthesis

Supervisor

Supervisor Agent
    ├── Worker 1
    ├── Worker 2
    └── Worker 3

Supervisor delegates and collects.
Use for: Complex features

Specialist Ensemble

Security Agent ─┐
Performance Agent ─┼── Decision Agent
Style Agent ─────┘

Multiple perspectives, one decision.
Use for: Code review, architecture

Dependency Management

Sequential Dependencies

Can't parallelize:
1. Create model
2. Create view using model
3. Create tests for view

Run in order. No agents.

Parallel Independence

Can parallelize:
1. Search auth/
2. Search payment/
3. Search user/

No dependencies. Spawn all.

Partial Dependencies

[A, B, C] → D → [E, F]

A, B, C parallel
D waits for all
E, F parallel after D

Handoff Protocol

Agent to Agent

Agent 1 completes:
"Found issue in auth.swift:142"

You relay to Agent 2:
"Fix the issue in auth.swift:142"

Context transferred explicitly.

Agent to Main

Agent returns summary.
You decide next action.
Main conversation has full control.

Background Check-In

run_in_background: true

Later:
TaskOutput to check status
Resume if needed

Complex Workflow Example

Feature Implementation

Phase 1 (Parallel):
- Design agent: Plan the feature
- Research agent: Find similar patterns

Phase 2 (After Phase 1):
- Implementation agent: Build it

Phase 3 (Parallel):
- Test agent: Write tests
- Docs agent: Write documentation

Phase 4:
- Review agent: Final check

Bug Investigation

Phase 1 (Parallel):
- Reproduce agent: Confirm the bug
- Search agent: Find related code

Phase 2:
- Analysis agent: Root cause

Phase 3:
- Fix agent: Implement solution

Phase 4 (Parallel):
- Test agent: Verify fix
- Regression agent: Check side effects

Coordination Commands

Start Workflow

"Start [workflow] for [target]"

Spawns appropriate agents.
Tracks progress.

Check Progress

"Status of [workflow]?"

Shows:
- Completed steps
- In progress
- Blocked
- Next steps

Adjust Workflow

"Skip [step]"
"Add [step] before [step]"
"Retry [step]"

Error Handling

Agent Fails

Agent hits error →
Report to main →
Decide: retry, skip, or abort

Dependency Blocked

Blocker identified →
Create workaround or
Escalate for decision

Timeout

Background agent too long →
Check with TaskOutput →
Decide: wait, cancel, or resume

Best Practices

✓ Do

- Clear handoff messages
- Explicit dependencies
- Background for long work
- Check-in periodically
- Handle errors gracefully

❌ Don't

- Implicit dependencies
- Fire and forget
- Over-orchestrate simple tasks
- Ignore failures

Use when: Multi-step features, complex investigations, coordinated work