Sub-Agent Orchestration Template
This is a reference template for creating skills that use parallel sub-agents. It is NOT directly invocable - use it as a pattern when building new orchestrated skills.
When to Use This Pattern
Use sub-agent orchestration when:
- Task has 3+ independent research phases
- Each phase can run in parallel without dependencies
- Results need synthesis into a unified output
- Context isolation improves quality (prevents contamination)
Do NOT use when:
- Task is simple and linear
- Phases depend on each other sequentially
- Interactive dialogue is needed throughout
Orchestration Structure
Phase 1: Planning
1. TaskCreate for overall work item
2. Identify independent workstreams
3. Create sub-tasks for each workstream
4. Set dependencies with TaskUpdate (addBlockedBy)
Phase 2: Parallel Execution
Launch N agents in parallel using Task tool:
Task tool calls (in single message for parallelism):
Agent 1: subagent_type: "Explore", model: "haiku"
- Purpose: [specific research goal]
- Tools: Read, Glob, Grep (read-only)
Agent 2: subagent_type: "Explore", model: "haiku"
- Purpose: [different research goal]
- Tools: Read, Glob, Grep (read-only)
Agent 3: subagent_type: "general-purpose"
- Purpose: [complex analysis requiring writes]
- Tools: All available
Phase 3: Synthesis
1. Collect results from all agents
2. Identify conflicts or gaps
3. Generate unified output
4. Update tasks to completed
Model Selection for Sub-Agents
| Sub-Agent Purpose | Model | Rationale |
|---|---|---|
| Read-only search/exploration | Haiku | Fast, cheap, isolated context |
| Analysis with context | Sonnet | Good reasoning |
| Complex synthesis | Sonnet | Balanced cost/quality |
| Deep architecture reasoning | Opus | Extended thinking needed |
Example: Parallel Research Pattern
## Phase 1: Create Tasks
TaskCreate: "Research API patterns"
TaskCreate: "Research database patterns"
TaskCreate: "Research auth patterns"
TaskCreate: "Synthesise findings" (blockedBy: above three)
## Phase 2: Launch Parallel Agents
[Single message with 3 Task tool calls]
Task 1: Explore agent for API patterns
Task 2: Explore agent for database patterns
Task 3: Explore agent for auth patterns
## Phase 3: Synthesise
Read agent outputs
Combine into unified report
Mark all tasks completed
Best Practices
- Limit to 3-5 parallel agents - More causes coordination overhead
- Use Haiku for exploration - Cheaper and faster for read-only
- Clear agent prompts - Each agent needs complete context (no shared state)
- Aggregate carefully - Agents may have conflicting findings
- Track with Tasks - Use TaskCreate/TaskUpdate for visibility
- Handle failures - If one agent fails, others continue
Anti-Patterns
❌ Don't spawn agents for sequential work ❌ Don't use background agents when MCP tools needed ❌ Don't pass partial context expecting shared state ❌ Don't spawn 10+ agents (diminishing returns)
Skills Using This Pattern
Reference these for working examples:
/vault-maintenance- Parallel health checks/quality-report- Parallel analysis dimensions/project-status- Parallel data gathering agents
Template Variables
When creating a new orchestrated skill, replace:
{{SKILL_NAME}} - Name of the skill
{{PHASE_COUNT}} - Number of parallel phases
{{AGENT_PURPOSE_N}} - Purpose of each agent
{{SYNTHESIS_OUTPUT}} - What the final output looks like
Source: DavidROliverBA/ArchitectKB — distributed by TomeVault.