Multi Agent Orchestration

Multi-Agent Orchestration Skill

Route and delegate tasks to the most appropriate AI provider based on task characteristics and provider capabilities.

Variables

Instructions

MANDATORY - Follow the Workflow steps below in order. Do not skip steps.

• Before delegating, understand the task characteristics
• Use the model-discovery skill for current model names
• Check cost/usage status before high-volume delegation

Quick Decision Tree

What type of task is this?
│
├─ Needs conversation history? ─────────► Keep in Claude (no delegation)
│
├─ Needs sandboxed execution? ──────────► OpenAI/Codex
│
├─ Large context (>100k tokens)? ───────► Gemini
│
├─ Multimodal (images/video)? ──────────► Gemini
│
├─ Needs web search? ───────────────────► Gemini
│
├─ Quick IDE edit? ─────────────────────► Cursor
│
├─ Privacy required / offline? ─────────► Ollama
│
├─ Provider-agnostic fallback? ─────────► OpenCode
│
└─ General reasoning / coding? ─────────► Claude (default)

Red Flags - STOP and Reconsider

If you're about to:

• Delegate without checking provider availability
• Use hardcoded model names (use model-discovery skill instead)
• Send sensitive data to a provider without user consent
• Delegate a task that requires your conversation history
• Skip the routing decision and guess which provider

STOP -> Read the appropriate cookbook file -> Check provider status -> Then proceed

Workflow

1. Analyze the task: What capabilities are required?
2. CHECKPOINT: Consult reference/provider-matrix.md for routing decision
3. Check provider availability: Run provider-check and cost-status if CHECK_COST_STATUS is true
4. Read the appropriate cookbook file for the selected provider
5. CHECKPOINT: Confirm API key / auth is configured
6. Execute delegation with proper context
7. Parse and summarize results for the user

Cookbook

Claude Code (Orchestrator)

• IF: Task requires complex reasoning, multi-file analysis, or conversation history
• THEN: Keep task in Claude Code (you are the orchestrator)
• WHY: Best for architecture decisions, complex refactoring

OpenAI / Codex

• IF: Task needs sandboxed execution OR security-sensitive operations
• THEN: Read and execute cookbook/openai-codex.md
• REQUIRES: OPENAI_API_KEY or Codex subscription

Google Gemini

• IF: Task involves large context (>100k tokens), multimodal (images/video), OR web search
• THEN: Read and execute cookbook/gemini-cli.md
• REQUIRES: GEMINI_API_KEY or Gemini subscription

Cursor

• IF: Task is quick IDE edits, simple codegen, or rename/refactor
• THEN: Read and execute cookbook/cursor-agent.md
• REQUIRES: Cursor installed and configured

OpenCode

• IF: Need provider-agnostic execution or a fallback CLI
• THEN: Read and execute cookbook/opencode-cli.md
• REQUIRES: OpenCode CLI installed and configured

Ollama (Local)

• IF: Task needs privacy, offline operation, or cost-free inference
• THEN: Read and execute cookbook/ollama-local.md
• REQUIRES: Ollama running with models pulled

Model Names

Do not hardcode model version numbers - they become stale quickly.

For current model names, use the model-discovery skill:

python .claude/ai-dev-kit/skills/model-discovery/scripts/fetch_models.py

Or read: .claude/ai-dev-kit/skills/model-discovery/SKILL.md

Quick Reference

See reference/provider-matrix.md for detailed routing guidance.

Tool Discovery

Orchestration tools are available in .claude/ai-dev-kit/dev-tools/orchestration/:

# Check provider status and usage
.claude/ai-dev-kit/dev-tools/orchestration/monitoring/cost-status.sh

# Check CLI availability (optional apply)
.claude/ai-dev-kit/dev-tools/orchestration/monitoring/provider-check.py

# Intelligent task routing
.claude/ai-dev-kit/dev-tools/orchestration/routing/route-task.py "your task"

# Direct provider execution
.claude/ai-dev-kit/dev-tools/orchestration/providers/claude-code/spawn.sh "task"
.claude/ai-dev-kit/dev-tools/orchestration/providers/codex/execute.sh "task"
.claude/ai-dev-kit/dev-tools/orchestration/providers/gemini/query.sh "task"
.claude/ai-dev-kit/dev-tools/orchestration/providers/cursor/agent.sh "task"
.claude/ai-dev-kit/dev-tools/orchestration/providers/opencode/execute.sh "task"
.claude/ai-dev-kit/dev-tools/orchestration/providers/ollama/query.sh "task"

Output

Delegation results should be:

1. Parsed from provider's response format
2. Summarized for the user
3. Integrated back into the conversation context

## Delegation Result

**Provider**: [provider name]
**Task**: [brief description]
**Status**: Success / Partial / Failed

### Summary
[Key findings or outputs]

### Details
[Full response if relevant]