Orchestrator

Orchestrator

Decompose complex work into independent waves of parallel sub-agent tasks.

When to Use

• Tasks with 3+ independent sub-tasks
• Research across multiple sources
• Bulk operations (triage, analysis, review)
• Any work where parallelism saves time

Procedure

1. Decompose Work

   • Break task into atomic, independent units
   • Group into waves (units within a wave have zero dependencies)
   • Order waves by dependency: wave 2 depends on wave 1 results

2. Plan Waves

   Wave 1 (parallel): [task-a, task-b, task-c]
   Wave 2 (parallel, depends on wave 1): [task-d, task-e]
   Wave 3 (sequential): [merge results]
   

3. Launch Sub-Agents

   • One Task agent per wave item
   • Each agent gets: clear objective, input data, output format
   • Each agent writes results to .claude/manifests/{task-id}.md

4. Collect Results

   • Wait for all agents in a wave to complete
   • Read manifests using manifest-reader skill
   • Check for failures or partial results

5. Merge and Synthesize

   • Combine results from all waves
   • Resolve conflicts or inconsistencies
   • Produce unified output

Wave Design Rules

• Max 5 agents per wave (resource constraint)
• Each agent should complete in under 2 minutes
• If an agent fails, note failure and continue (don't block the wave)
• Always have a merge step as the final wave

Manifest Format

Each sub-agent writes:

---
task: {task-id}
status: complete | failed | partial
agent: {agent-description}
---

## Findings
{results}

## Errors
{any errors encountered}

Output Format

Orchestration: {task name}
Waves: {count}
Results: {success count}/{total count} tasks completed
Summary: {merged findings}
Failures: {list or "none"}

Rules

• Never launch wave 2 before wave 1 completes
• Always check manifests before proceeding
• Failed agents should not block successful ones
• Keep each agent's scope small and focused
• Clean up manifest files after final merge