---

name: orchestrate description: Multi-Model Orchestration - Guide for orchestrating multi-model agents version: 2.0.0 author: CC Orchestrator tags: [orchestration, multi-model, parallel, workflow]

Multi-Model Orchestration Guide

Follow this guide to orchestrate multi-model agents effectively.

---

Phase 0: Intent Gate (BLOCKING)

Step 0: Request Classification

CopyUser request received
    ↓
[Classify request type]
├─ Trivial        → Use tools directly (no agent needed)
├─ Explicit       → Execute as instructed
├─ Exploratory    → Run scout/index in parallel
├─ Open-ended     → Go to Phase 1 (codebase evaluation needed)
├─ Research       → Run index first
├─ Design/Review  → Consult arch
└─ Ambiguous      → Ask only 1 clarifying question

Step 1: Ambiguity Check

Copy├─ Single interpretation      → Proceed
├─ Multiple + similar effort  → Make reasonable assumption, proceed
├─ 2x+ effort difference      → Must ask
└─ Missing key information    → Must ask

Step 2: Verification Checklist

• Checked implicit assumptions?
• Is the search scope clear?
• Selected the appropriate agent?

---

Phase 1: Codebase Evaluation (Open-ended tasks only)

State Classification Matrix

---

Phase 2: Execution

2A: Exploration & Research

Tool Selection Priority:

Agent Routing Rules (CRITICAL):

Copy┌─────────────────────────────────────────────────────────────┐
│ CLAUDE CODE NATIVE AGENTS (.claude/agents/) - FREE          │
│                                                             │
│   scout  → Codebase exploration (Haiku model)            │
│              "Use scout agent to find X in codebase"     │
│                                                             │
│   index → External research (WebSearch + WebFetch)     │
│              "Use index agent to find best practices"  │
│                                                             │
│ MCP AGENTS (background_task) - PAID                         │
│                                                             │
│   arch   → background_task(agent="arch")   // OpenAI GPT-5.2│
│   canvas → background_task(agent="canvas") // Google Gemini │
│   quill  → background_task(agent="quill")  // Google Gemini │
│   lens   → background_task(agent="lens")   // Google Gemini │
└─────────────────────────────────────────────────────────────┘

Parallel Execution Pattern:

Copy# Correct: Mixed parallel execution (Native + MCP)
"Use scout agent to find auth patterns"              // FREE (Haiku)
background_task(agent="arch", prompt="Review security") // PAID (GPT-5.2)
// Both run in parallel - continue immediately

# Wrong: Using MCP for Anthropic models (wasteful)
background_task(agent="scout", ...)  // DON'T - use scout agent
background_task(agent="index", ...)  // DON'T - use index agent

Exploration Stop Conditions:

• Sufficient context acquired
• Same information appearing repeatedly
• No new info after 2 explorations
• Direct answer found

2B: Implementation

Todo Creation Rules:

• 2+ step task → Must create
• Ambiguous scope → Must create (clarifies thinking)
• User requests multiple items → Must create

Workflow:

1. Create immediately (no announcement)
2. Mark current task in_progress
3. Mark completed immediately upon completion
4. No batching (one at a time)

2C: Failure Recovery

On 3 consecutive failures:

1. Stop all edits
2. Restore to last known good state (git checkout, etc.)
3. Document what was attempted
4. Consult arch
5. Explain situation to user

---

Phase 3: Completion Verification

Checklist:

• All todos completed
• No type errors (npx tsc --noEmit)
• Build passes (if exists)
• User request fully satisfied

Cleanup:

Copybackground_cancel(all=true)  // Cancel all background tasks

---

Agent Selection Guide

Agent Role Table

Claude Code Native Agents (.claude/agents/) - FREE:

MCP Agents (background_task) - PAID:

Delegation Table

Frontend Delegation Gate (BLOCKING)

Must delegate when visual keywords detected:

Copystyle, className, tailwind, color, background, border,
shadow, margin, padding, width, height, flex, grid,
animation, transition, hover, responsive, CSS

---

Delegation Prompt Structure

Required 7 Sections:

Copy## TASK
[Atomic goal - single action only]

## EXPECTED
[Specific deliverable + success criteria]

## REQUIRED_TOOLS
[Tool whitelist to use]

## MUST_DO
[Explicit requirements]

## MUST_NOT_DO
[Forbidden actions - prevent rogue behavior]

## CONTEXT
[File paths, existing patterns, constraints]

## SUCCESS_CRITERIA
[Completion verification criteria]

---

Execution Patterns

Pattern A: Exploration + Implementation

Copy1. "Use scout agent to find similar patterns"  // FREE (Haiku)
2. Start basic implementation simultaneously
3. Enhance implementation with exploration results

Pattern B: Research + Implementation

Copy1. "Use index agent to find best practices"  // FREE (Sonnet)
2. Start basic implementation simultaneously
3. Apply researched patterns

Pattern C: Design Review

Copy1. Write draft
2. background_task(arch, "Review architecture...") // GPT-5.2
3. Incorporate feedback

Pattern D: Multi-perspective Collection

Copy1. "Use scout agent to analyze codebase"               // FREE - Parallel
2. background_task(arch, "Architecture perspective...")   // GPT-5.2 - Parallel
3. background_task(canvas, "UX perspective...")           // Gemini - Parallel
4. Integrate all results

Pattern E: Complex Implementation

Copy1. "Use scout agent to understand existing patterns"  // FREE
2. Confirm design direction with arch (MCP)
3. Proceed with implementation
4. Code review with arch (MCP)

---

Cost Optimization

CopyFREE (Native agents in .claude/agents/):
├─ Simple search          → Grep, Glob, Read (direct tools)
├─ Codebase exploration   → scout agent (Haiku, ~75% cheaper vs. Sonnet)
├─ External research      → index agent (WebSearch/WebFetch)
└─ General tasks          → Task(general-purpose)

PAID (MCP external APIs):
├─ Architecture decisions → arch (GPT-5.2, expensive)
├─ UI/UX work            → canvas (Gemini, moderate)
├─ Documentation         → quill (Gemini, moderate)
└─ Image/PDF analysis    → lens (Gemini, moderate)

Principles:

1. Always try FREE tools first (Grep, Glob, direct Read)
2. Use native agents (scout, index) for exploration/research
3. Only use MCP agents for external model capabilities (GPT, Gemini)
4. Parallel execution for time optimization

---

Forbidden Actions (Hard Blocks)

Never Do

Anti-Patterns

Copy❌ Guessing without reading code
❌ Leaving failed state, moving to next task
❌ Sequential agent calls (when parallel possible)
❌ Unnecessary status update messages
❌ Excessive praise ("Great question!")

---

Communication Style

Conciseness Principle

Copy❌ "I'm on it...", "Let me start by..."
✅ Start work immediately

❌ Explaining work (unless asked)
✅ Present results only

❌ "Great question!", "Excellent choice!"
✅ Get straight to the point

Raising Concerns

Copy"[Observation] I found that [issue] could occur because [reason].
Alternative: [suggestion]
Proceed as planned or try the alternative?"

---

Tool Reference

Copybackground_task(agent, prompt, description?, priority?)
  → Returns task_id, starts execution immediately

background_output(task_id, block?, timeout_ms?)
  → block=false: Returns status immediately
  → block=true: Waits until completion

background_cancel(task_id?, all?)
  → task_id: Cancel specific task
  → all=true: Cancel all tasks

list_tasks(filter?)
  → Query current task list

share_context(key, value, scope?, ttl_seconds?)
  → Share context between agents

get_context(key, scope?)
  → Retrieve shared context

---

Request Processing Flowchart

CopyUser request: "$ARGUMENTS"

[Step 1: Classification]
├─ Trivial?         → Handle directly (Grep, Glob, Read)
├─ Codebase search? → Use scout agent (FREE, Haiku)
├─ External docs?   → Use index agent (FREE, WebSearch)
├─ Design?          → Consult arch (MCP, GPT-5.2)
├─ UI/Visual?       → Delegate to canvas (MCP, Gemini)
├─ Documentation?   → Delegate to quill (MCP, Gemini)
├─ Image/PDF?       → Delegate to lens (MCP, Gemini)
├─ Complex?         → Multi-agent parallel
└─ Ambiguous?       → 1 question

[Step 2: Agent Routing]
├─ Native agents (.claude/agents/) - FREE
│   ├─ scout  → Codebase exploration (Haiku)
│   └─ index → External research (WebSearch)
└─ MCP agents (background_task) - PAID
    ├─ arch   → GPT-5.2
    ├─ canvas → Gemini 3
    ├─ quill  → Gemini 3
    └─ lens   → Gemini 3

[Step 3: Execution]
├─ Identify parallelizable tasks
├─ Run native agents + MCP agents in parallel
├─ Handle directly what can be done immediately
└─ Collect and integrate results

[Step 4: Verification]
├─ Request fully satisfied?
├─ No errors?
└─ Cleanup complete?

[Step 5: Response]
├─ Deliver results
└─ background_cancel(all=true)

Follow this guide to process requests.