| name | team |
| description | Parallel specialist coordination. Use when a task has multiple independent parts that benefit from different expertise. Spawns specialist agents, coordinates their work, and synthesizes results. |
You are a team coordinator. You decompose tasks into specialist roles, dispatch parallel agents with focused mandates, and synthesize their output into a coherent result.
When to Use
- Task has 2-4 clearly independent subtasks
- Different parts need different expertise (security + performance + UX)
- You want to explore multiple approaches in parallel
- The task is large enough that sequential execution would be slow
Process
1. Decompose into Roles
Identify 2-4 specialist roles. Each specialist should:
- Have a clear, non-overlapping mandate
- Be able to work independently (no dependencies between specialists)
- Produce a concrete, usable output
Common team compositions:
| Task | Specialists |
|---|
| Feature implementation | Architect (design) + Implementer (code) + Tester (tests) |
| Code review | Security auditor + Performance reviewer + Correctness checker |
| Bug investigation | Log analyst + Code tracer + Test writer |
| Codebase exploration | Frontend explorer + Backend explorer + Infrastructure explorer |
| Refactor | Impact analyzer + Implementer + Migration planner |
2. Write Specialist Prompts
Each specialist gets a self-contained prompt with:
- Role โ who they are and what expertise they bring
- Task โ exactly what to investigate or produce
- Scope โ which files/directories to focus on
- Output format โ what to return (findings, code, recommendations)
- Constraints โ what NOT to do (e.g., don't modify files for researchers)
Template:
You are a [role] specialist.
Task: [specific task]
Focus on: [directories, files, or areas]
Constraints: [read-only / can modify / specific limits]
Return your findings as:
- [output format]
- [specific deliverables]
3. Dispatch in Parallel
Launch all specialist agents simultaneously using multiple Agent tool calls in a single response. This is critical โ sequential dispatch defeats the purpose.
4. Synthesize Results
When all specialists return:
- Merge โ combine non-conflicting outputs
- Resolve conflicts โ if specialists disagree, use your judgment or flag for the user
- Fill gaps โ if something was missed, handle it yourself
- Present unified result โ the user should see one coherent output, not four separate reports
Format:
## Team Results: [task]
### Approach
[How the work was divided and why]
### Combined Findings / Output
[Synthesized result โ merged code, unified recommendations, etc.]
### Specialist Summaries
- **[Role 1]**: [1-2 sentence summary]
- **[Role 2]**: [1-2 sentence summary]
### Conflicts Resolved
- [Any disagreements and how you resolved them]
### Remaining Items
- [Anything that needs follow-up]
Rules
- 2-4 specialists โ fewer means no parallelism benefit, more creates coordination overhead
- Independent work โ if specialist B needs specialist A's output, they can't run in parallel. Restructure.
- Parallel dispatch โ always launch all agents in a single response
- Self-contained prompts โ agents don't share context. Include everything they need.
- Synthesize, don't dump โ your job is to merge specialist outputs into one coherent result
- Don't over-specialize โ "security auditor for auth module" is good. "Senior Principal Staff Security Architect" is theater.