Multi-Agent Orchestrator

---
name: multi-agent-orchestrator
description: Gives an agent the ability to coordinate multiple specialist agents using the orchestrator-workers pattern: decompose tasks, delegate, monitor progress, and synthesize results.
inputs:
  - task: Complex task description needing multiple capabilities
  - agent_registry: List of available specialist agents with their capabilities
  - delegation_rules: How to assign subtasks (by capability, load, priority)
  - synthesis_format: How to combine worker outputs (merge, evaluate, rank)
outputs:
  - orchestration_plan: Decomposed subtasks with agent assignments
  - execution_log: Per-worker status, outputs, and error handling
  - synthesized_result: Combined final output from all workers
  - metrics: Task completion time, agent utilization, success rate
tools:
  - openai_claude: Orchestrator LLM for task decomposition and synthesis
  - task_queue_postgres: PostgreSQL-backed queue for subtask tracking
  - squish_memory: Cross-session agent memory and context passing
  - mcp_tools: Agent tool integrations for specialist capabilities
safety:
  - Set maximum iterations and timeouts per worker agent
  - Log all delegation decisions for audit
  - Never delegate sensitive actions (deployments, payments) without human approval
  - Implement circuit breaker: if a worker fails 3 times, route to human
  - Validate worker outputs before synthesis to catch hallucinations
---

# Multi-Agent Orchestrator Skill

Coordinate multiple specialist agents using the orchestrator-workers pattern: decompose complex tasks, delegate to specialists, monitor progress, and synthesize results.

## When to Use

- You have a task that requires multiple capabilities (research + write + review)
- You want to parallelize work across specialist agents
- You need a central coordinator to track complex multi-step workflows
- You are building the Internal AI OS and need the orchestration layer

## How It Works

1. **Decompose**: Orchestrator analyzes task and breaks into subtasks with dependencies
2. **Assign**: Each subtask is assigned to the best-fit agent from the registry
3. **Execute**: Workers run in parallel or sequence based on dependency graph
4. **Monitor**: Orchestrator tracks progress, handles failures, re-routes as needed
5. **Synthesize**: Combine worker outputs into coherent final result

## Orchestration Patterns

- **Fan-out**: One task to many workers in parallel (best for research/data gathering)
- **Pipeline**: Sequential subtasks where each depends on previous output (best for content creation)
- **Hybrid**: Parallel groups with sequential dependencies (best for complex projects)
- **Competitive**: Multiple workers on same task, pick best result (best for creative work)

## Example Prompt

"Set up a multi-agent system for writing a blog post. Use three specialist agents: 1) ResearchAgent - finds latest stats and examples on the topic, 2) WriterAgent - drafts the post based on research, 3) EditorAgent - reviews and polishes. Orchestrate them in a pipeline. Use Squish memory so WriterAgent can access ResearchAgent's findings."

## Related

- Recipe: /recipes/internal-ai-os
- Recipe: /recipes/agent-memory-squish