Agent Orchestration Patterns

Agent Orchestration Patterns

Build coordinated AI teams that work together seamlessly

This skill provides battle-tested patterns for orchestrating multiple AI agents to solve complex problems collaboratively.

Core Principles

1. Single Responsibility Agents

Each agent should have one clear purpose. Generalists create confusion; specialists create excellence.

Bad:
  GeneralAgent: "Does everything - UI, backend, writing, testing"

Good:
  FrontendAgent: "React components, styling, accessibility"
  BackendAgent: "APIs, database, business logic"
  TestAgent: "Unit tests, integration tests, E2E"

2. Clear Communication Protocols

Define how agents share information, hand off work, and resolve conflicts.

Handoff Protocol:
  From: BackendAgent
  To: FrontendAgent
  Includes:
    - API contract (types, endpoints)
    - Example payloads
    - Error scenarios
    - Authentication requirements

3. Orchestrator Pattern

One agent coordinates; others execute. Prevents chaos and conflicting directions.

                    ┌─────────────────┐
                    │   ORCHESTRATOR  │
                    │  (Coordinates)  │
                    └────────┬────────┘
                             │
        ┌────────────────────┼────────────────────┐
        │                    │                    │
        ▼                    ▼                    ▼
┌───────────────┐   ┌───────────────┐   ┌───────────────┐
│   Agent A     │   │   Agent B     │   │   Agent C     │
│  (Executes)   │   │  (Executes)   │   │  (Executes)   │
└───────────────┘   └───────────────┘   └───────────────┘

Orchestration Patterns

Pattern 1: Sequential Pipeline

Best for: Tasks with clear dependencies where each step requires the previous step's output.

Pipeline:
  1. Research Agent → Gathers information
  2. Analysis Agent → Processes findings (needs step 1)
  3. Writing Agent → Creates content (needs step 2)
  4. Review Agent → Quality check (needs step 3)

Implementation:

## Sequential Pipeline Protocol

### Step 1: Research Phase
**Agent:** Research Agent
**Input:** User query
**Output:** Structured research document
**Completion Signal:** "Research complete. Findings ready for analysis."

### Step 2: Analysis Phase
**Agent:** Analysis Agent
**Input:** Research document from Step 1
**Output:** Analyzed insights with recommendations
**Completion Signal:** "Analysis complete. Ready for content creation."

### Step 3: Writing Phase
**Agent:** Writing Agent
**Input:** Analysis from Step 2
**Output:** Draft content
**Completion Signal:** "Draft complete. Ready for review."

### Step 4: Review Phase
**Agent:** Review Agent
**Input:** Draft from Step 3
**Output:** Final content with quality assessment
**Completion Signal:** "Review complete. Content finalized."

Pattern 2: Parallel Fan-Out

Best for: Independent tasks that can run simultaneously.

Fan-Out:
  Orchestrator splits task into:
    - Agent A: Frontend components
    - Agent B: Backend APIs
    - Agent C: Database schema

  Fan-In:
    - Orchestrator collects results
    - Integrates into unified solution

Implementation:

## Parallel Fan-Out Protocol

### Split Phase
**Orchestrator Action:** Divide task into independent workstreams
**Criteria for parallelization:**
- No shared state dependencies
- No sequential requirements
- Clear interface contracts defined

### Parallel Execution
**Agent A:** [Task description]
- Works independently
- Reports completion status

**Agent B:** [Task description]
- Works independently
- Reports completion status

**Agent C:** [Task description]
- Works independently
- Reports completion status

### Integration Phase
**Orchestrator Action:**
- Wait for all agents to complete
- Resolve any interface conflicts
- Integrate outputs into unified result

Pattern 3: Specialist Consultation

Best for: Tasks requiring domain expertise at specific points.

Consultation:
  Primary Agent working...
  → Hits domain-specific challenge
  → Consults Specialist Agent
  → Receives expert guidance
  → Continues with enhanced solution

Implementation:

## Specialist Consultation Protocol

### Recognition Triggers
The primary agent should consult a specialist when:
- Task requires domain-specific knowledge
- Decision has significant architectural impact
- Quality standard requires expert validation
- Risk mitigation requires specialized review

### Consultation Format
**From:** [Primary Agent]
**To:** [Specialist Agent]
**Context:** [What we're building]
**Question:** [Specific question]
**Constraints:** [Relevant limitations]
**Expected Output:** [What we need back]

### Response Integration
Specialist provides:
- Direct answer to question
- Reasoning behind recommendation
- Potential alternatives considered
- Caveats or edge cases

Pattern 4: Debate & Synthesis

Best for: Complex decisions where multiple perspectives improve outcomes.

Debate:
  Agent A: Argues for Approach 1
  Agent B: Argues for Approach 2
  Agent C: Synthesizes best of both
  Orchestrator: Makes final decision

Implementation:

## Debate & Synthesis Protocol

### Phase 1: Position Development
**Agent A:** Develops Position 1
- State the approach clearly
- List all advantages
- Acknowledge weaknesses
- Provide implementation path

**Agent B:** Develops Position 2
- State the approach clearly
- List all advantages
- Acknowledge weaknesses
- Provide implementation path

### Phase 2: Critique
Each agent critiques the other's position:
- What's missing?
- What's overestimated?
- What edge cases are unhandled?

### Phase 3: Synthesis
**Synthesis Agent:**
- Extract best elements from each position
- Resolve contradictions
- Propose hybrid solution if beneficial

### Phase 4: Decision
**Orchestrator:**
- Evaluate all positions and synthesis
- Make final decision with reasoning
- Document decision rationale

Pattern 5: Hierarchical Delegation

Best for: Large projects requiring multiple levels of coordination.

                      ┌─────────────────┐
                      │  ORCHESTRATOR   │
                      │   (Strategic)   │
                      └────────┬────────┘
                               │
           ┌───────────────────┼───────────────────┐
           │                   │                   │
           ▼                   ▼                   ▼
   ┌───────────────┐   ┌───────────────┐   ┌───────────────┐
   │  Team Lead A  │   │  Team Lead B  │   │  Team Lead C  │
   │  (Tactical)   │   │  (Tactical)   │   │  (Tactical)   │
   └───────┬───────┘   └───────┬───────┘   └───────┬───────┘
           │                   │                   │
    ┌──────┴──────┐     ┌──────┴──────┐     ┌──────┴──────┐
    │             │     │             │     │             │
    ▼             ▼     ▼             ▼     ▼             ▼
┌───────┐   ┌───────┐   ...         ...   ...         ...
│Agent 1│   │Agent 2│
└───────┘   └───────┘

Communication Contracts

Agent-to-Agent Message Format

Message:
  from: "agent_id"
  to: "agent_id"
  type: "request|response|status|handoff"
  priority: "low|normal|high|critical"
  content:
    summary: "Brief description"
    details: "Full content"
    artifacts: ["list of outputs"]
  context:
    conversation_id: "unique_id"
    parent_message: "optional_id"
    related_tasks: ["task_ids"]

Status Reporting Protocol

Status Update:
  agent: "agent_id"
  timestamp: "ISO 8601"
  status: "idle|working|blocked|complete|error"
  current_task: "description"
  progress: "0-100%"
  blockers: ["list of blockers"]
  next_steps: ["planned actions"]
  estimated_completion: "optional timestamp"

Error Escalation Protocol

Error Report:
  agent: "agent_id"
  severity: "warning|error|critical"
  error_type: "category"
  description: "what went wrong"
  attempted_solutions: ["what we tried"]
  suggested_actions: ["what might help"]
  requires_human: true|false
  blocks_progress: true|false

Quality Gates

Before Agent Assignment

• Task clearly defined
• Success criteria established
• Dependencies mapped
• Appropriate agent selected
• Resources available

During Execution

• Progress being reported
• Blockers escalated promptly
• Quality standards maintained
• Timeline on track

After Completion

• Deliverables meet criteria
• No known defects
• Documentation complete
• Handoff information ready

Anti-Patterns to Avoid

1. The Generalist Trap

Problem: One agent tries to do everything Symptom: Inconsistent quality, context overload Solution: Split into specialized agents

2. The Circular Dependency

Problem: Agent A waits for B, B waits for C, C waits for A Symptom: Deadlock, no progress Solution: Identify and break cycles, define clear ordering

3. The Silent Agent

Problem: Agent works without status updates Symptom: Orchestrator has no visibility, surprises at completion Solution: Require regular status reports

4. The Micro-Manager

Problem: Orchestrator controls every small decision Symptom: Bottleneck at orchestrator, slow progress Solution: Delegate decisions within boundaries

5. The Scope Creeper

Problem: Agent expands task beyond assignment Symptom: Delayed completion, unnecessary work Solution: Clear scope definition, confirmation before expansion

Team Templates

Minimal Development Team (3 agents)

Team:
  Architect:
    Role: Orchestrator + Technical decisions
    Responsibilities: Planning, coordination, architecture

  Builder:
    Role: Implementation
    Responsibilities: Frontend, backend, integrations

  Validator:
    Role: Quality assurance
    Responsibilities: Testing, review, documentation

Standard Development Team (5 agents)

Team:
  Architect:
    Role: Orchestrator

  Frontend:
    Role: UI specialist

  Backend:
    Role: API/data specialist

  DevOps:
    Role: Infrastructure

  QA:
    Role: Testing specialist

Full Product Team (8+ agents)

Team:
  Product:
    Strategist: Vision and roadmap
    Designer: UX/UI design

  Engineering:
    Architect: Technical leadership
    Frontend: UI implementation
    Backend: Services implementation
    DevOps: Infrastructure
    QA: Testing

  Content:
    Writer: Documentation, copy

Integration with Claude Code

Agent Definition Format

# .claude/agents/example-agent.md

---
name: Example Agent
description: What this agent does
model: sonnet|opus
mcpServers:
  - server-name
workingDirectories:
  - /path/to/focus
---

# Agent Name

## Mission
What this agent aims to accomplish.

## Responsibilities
- Specific duty 1
- Specific duty 2

## Capabilities
- What tools/skills it can use

## Communication Protocol
How it reports status and coordinates.

Skill Loading

# Reference this skill in your session
Skills:
  - .claude/skills/community/agent-orchestration

"A team of specialized agents, well-coordinated, will always outperform a single generalist."