Multi-Agent Planner

Purpose

Design a multi-agent orchestration architecture: agent roles, communication patterns, handoff protocols, and failure recovery for collaborative AI systems.

Input: Use case description, AI architecture (optional), task decomposition, constraints (latency, cost, reliability)
Output: Multi-Agent Orchestration Plan with agent roster, interaction diagram, and coordination rules Examples: See examples.md for worked input/output.

Workflow

Step 1: Decompose the problem

Break the use case into subtasks:

Identify which subtasks require specialized reasoning, tool access, or isolation.

Step 2: Define agent roles

For each agent specify:

• Responsibility — single bounded objective
• Inputs — what context it receives
• Outputs — artifact or message format
• Tools — allowed integrations (MCP, APIs, none)
• Autonomy level — fully autonomous / supervised / advisory only

Step 3: Select orchestration pattern

Document chosen pattern with rationale.

Step 4: Design communication protocol

Define:

• Message schema (task ID, agent ID, payload, status)
• Handoff triggers (completion, escalation, timeout)
• Shared state store (conversation, scratchpad, blackboard)
• Conflict resolution (priority rules, supervisor override)

Step 5: Plan failure and recovery

Step 6: Validate

Run Validation checklist.

Decision Rules

Validation

• Problem decomposed with dependency map
• Each agent has single bounded responsibility
• Orchestration pattern selected with rationale
• Message schema and handoff triggers defined
• Failure modes with recovery actions documented
• Tool permissions scoped per agent
• Cost/latency impact estimated for agent count
• Human-in-the-loop gates defined for high-risk paths

Anti-patterns

• Agent sprawl — creating agents for every minor subtask without coordination benefit.
• Shared omniscient context — passing full conversation history to every agent.
• No termination condition — agents loop without max iteration or success criteria.
• Duplicate responsibilities — two agents performing the same task without clear split.
• Orchestration without eval — multi-agent system without per-agent quality measurement.

Best Practices

• Start with supervisor pattern; evolve to peer-to-peer only when justified.
• Keep agent count ≤5 for maintainability unless domain requires more.
• Use structured output contracts between agents (JSON schema, not free text).
• Log every inter-agent message for debugging and audit.
• Pair with ai-evaluation-builder to measure per-agent contribution.

Output Structure

# Multi-Agent Orchestration Plan: [System Name]

## Problem Decomposition
| Subtask | Agent candidate | Dependencies |
|---------|-----------------|--------------|

## Agent Roster
| Agent | Responsibility | Tools | Autonomy |
|-------|----------------|-------|----------|

## Orchestration Pattern
[Pattern name and rationale]

## Interaction Diagram
[ASCII or mermaid sequence]

## Communication Protocol
| Field | Type | Description |
|-------|------|-------------|

## Failure & Recovery
| Failure | Detection | Recovery |
|---------|-----------|----------|

## Cost & Latency Estimate
[Per-request impact vs single-agent baseline]

## Open Decisions
- [ ] [Unresolved design choice]

Next Skills