Multi-Agent Collaboration

Patterns for effective agent coordination.

Instructions

1. Agent Role Definition

Specialist Roles

2. Task Delegation

Delegation Protocol

1. Assess: What expertise is needed?
2. Assign: Match task to specialist
3. Brief: Provide clear context
4. Monitor: Track progress
5. Integrate: Combine results

Delegation Template

TASK: [Description] ASSIGNED TO: [Role] CONTEXT: [What they need to know] DELIVERABLE: [Expected output] DEADLINE: [When needed]

3. Communication Protocols

Message Format

Request

{ "from": "planner", "to": "developer", "type": "task", "content": { "action": "implement", "file": "UserService.ts", "requirements": ["..."] } }

Response

{ "from": "developer", "to": "planner", "type": "result", "status": "complete", "content": { "files_changed": ["UserService.ts"], "summary": "Added user validation" } }

4. Handoff Patterns

Clean Handoff Checklist

When passing work to another agent:

• Summarize what was done
• Explain current state
• List remaining tasks
• Note any blockers
• Provide relevant context

Example

Handoff: Planner → Developer

Completed

• Analyzed requirements

• Identified 3 main components

Current State

• Types defined in types/user.ts

• No implementation yet

Next Steps

• Create UserService in services/

• Implement CRUD operations

• Add validation

Context

• Use Prisma for DB

• Follow existing patterns in OrderService

5. Conflict Resolution

When Agents Disagree

1. Identify: What is the conflict?
2. Evidence: Gather supporting facts
3. Evaluate: Compare approaches
4. Decide: Choose best option
5. Document: Record decision and reasoning

Common Conflicts

6. Workflow Orchestration

sequenceDiagram participant User participant Planner participant Developer participant Reviewer

User->>Planner: Request feature
Planner->>Planner: Analyze & decompose
Planner->>Developer: Assign implementation
Developer->>Developer: Write code
Developer->>Reviewer: Request review
Reviewer->>Developer: Feedback
Developer->>Planner: Complete
Planner->>User: Deliver result

7. Shared Context

Context Sharing Rules

Must Share

• Project constraints
• User preferences
• Critical decisions
• Error states

May Share

• Implementation details
• Intermediate results
• Research findings

Don't Share

• Internal reasoning
• Failed attempts
• Verbose logs

8. Progress Synchronization

Status Updates

Format

Sync Points

• After each phase completion
• When blocked
• When priorities change
• Before final delivery

References

• Multi-Agent Reinforcement Learning

• CrewAI Framework

• AutoGen