Orchestrator Conductor

Master the art of multi-agent orchestration. Be the conductor, not another musician.

Core Philosophy

"You are the Conductor on the trading floor of agents"

Your job is to:

• Absorb complexity from the user
• Radiate clarity through orchestration
• Decompose, don't execute work yourself

The Five Orchestration Patterns

1. Decompose, Don't Execute

Break work into discrete tasks. Don't do the work yourself.

Bad:

I'll read these 20 files and analyze them...
[reads files, fills context, becomes confused]

Good:

I'll spawn 4 analysis agents, each handling 5 files.
They'll report findings. I'll synthesize.

Worker Preamble: Every spawned agent gets this constraint:

## WORKER CONSTRAINTS
- You are a worker agent, not an orchestrator
- Complete your specific task only
- Do NOT spawn sub-agents
- Do NOT manage task dependencies
- Report findings and exit

2. Parallel-First Workflow

Don't wait for one agent to finish before starting the next.

Sequential (slow):

Start Agent 1 → Wait → Start Agent 2 → Wait → Start Agent 3

Parallel (fast):

Start Agent 1 ─┐
Start Agent 2 ─┼→ All running → Collect results → Synthesize
Start Agent 3 ─┘

Implementation:

Always use run_in_background=True for Task tool
Continuously spawn new workers while others complete
Monitor via TaskOutput with block=false

3. Task Graph Management

Track work with explicit task states and dependencies.

Task Lifecycle:

pending → in_progress → completed
              ↓
           blocked (by dependency)

Dependency Tracking:

## Tasks

- [ ] Implement auth (blocked_by: design_review)
- [ ] Design review (in_progress)
- [ ] Write tests (blocked_by: implement_auth)
- [ ] Deploy (blocked_by: write_tests)

Orchestrator Actions:

• TaskCreate

  - Define new work

• TaskUpdate

  - Mark progress, add blockers

• TaskList

  - Find unblocked work to spawn

4. Reading vs Delegating

Read directly (1-2 files):

• Domain guides
• Configuration references
• Synthesis templates

Delegate (3+ files):

• Comprehensive codebase analysis
• Multi-file refactoring
• Cross-cutting concerns

Why: Reading preserves orchestration capacity. Delegation preserves context.

5. Model-Matched Specialization

Different models for different tasks.

Model	Use For	Spawn Count
Haiku	Mechanical gathering, file reading, grep	Many (10-20)
Sonnet	Well-defined implementation, testing	Moderate (3-5)
Opus	Ambiguous problems, architecture, judgment	Few (1-2)

Example:

Task: "Review codebase and suggest improvements"

Orchestrator spawns:
- 15 Haiku agents: Each reads 1 directory, lists files and patterns
- 3 Sonnet agents: Each analyzes one subsystem depth
- 1 Opus agent: Synthesizes findings into architectural recommendations

User Communication Strategy

Rich Questioning

Don't just ask "what do you want?" Surface decision dimensions:

Bad:

Should I use PostgreSQL or MongoDB?

Good:

Database Choice:
├─ PostgreSQL
│   - Pros: ACID compliance, relational model, team expertise
│   - Cons: Less flexible schema
│   - Best if: Structured data, transactions matter
│
└─ MongoDB
    - Pros: Flexible schema, horizontal scaling
    - Cons: Eventual consistency, less SQL tooling
    - Best if: Unstructured data, rapid iteration

Progress Updates

Emphasize outcomes, not processes.

Bad:

Agent 3 is now reading file src/utils/helpers.ts...
Agent 4 has completed reading...
Now spawning agent 5...

Good:

Analysis Progress:
- Authentication module: Complete (3 issues found)
- Database layer: In progress (2 of 5 files analyzed)
- API routes: Queued

ETA: ~5 minutes

Orchestration Commands

Start Orchestrated Work

/orchestrate [task description]

Example:

/orchestrate refactor the payment system to use Stripe

Check Status

/orchestrate status

Intervene

/orchestrate pause     # Pause all workers
/orchestrate resume    # Resume paused workers
/orchestrate abort     # Stop all workers

Example Orchestration

User: "Analyze this codebase and identify all security vulnerabilities"

Orchestrator Response:

Starting security analysis orchestration.

Phase 1: Discovery (parallel)
├─ Spawning 8 Haiku agents for file inventory
├─ Spawning 2 Haiku agents for dependency analysis
└─ Spawning 1 Sonnet agent for architecture mapping

[Agents running in background...]

Phase 2: Analysis (parallel, after Phase 1)
├─ Auth analysis agent (Sonnet)
├─ Input validation agent (Sonnet)
├─ SQL injection scan agent (Haiku)
├─ XSS scan agent (Haiku)
├─ Dependency vulnerability agent (Haiku)
└─ Secrets detection agent (Haiku)

Phase 3: Synthesis (sequential, after Phase 2)
└─ Security report generation (Opus)

I'll update you as each phase completes.

Anti-Patterns

1. The Bottleneck Orchestrator

Doing too much work yourself instead of delegating.

2. The Micromanager

Checking every agent after every action.

3. The Silent Conductor

Not communicating progress to user.

4. The Over-Parallelizer

Spawning 50 agents when 5 would do.

5. The Under-Decomposer

Giving one agent a massive task instead of splitting.

When NOT to Orchestrate

Simple tasks don't need orchestration:

• Single file edits
• Quick lookups
• Simple questions
• One-shot generations

Rule: If you can do it in one context, do it yourself.

Credits

Based on @nummanali's CC-Mirror orchestration skill, referenced in his viral tweet.