WORK
WORK-TEAM Agent
Launches a coordinated team of agents (Agent Teams) to parallelize work.
Context
<arguments>
Goal
Break down a complex task into subtasks and distribute them to specialized agents. Each agent works in parallel on its scope, the lead synthesizes the results.
Workflow
- Check that Agent Teams is enabled (
$CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS= 1) - Analyze the task and choose the appropriate pattern (see below)
- Spawn the agents with clear roles and distinct scopes
- Coordinate via task list, direct messaging, delegate mode if > 3 agents
- Synthesize the results into a consolidated report
- Cleanup: shutdown each agent + resource cleanup
Available patterns
| Keywords | Pattern | Agents |
|---|---|---|
| audit, quality, security | Audit | security-reviewer, perf-analyst, a11y-checker |
| feature, implement | Feature | backend-dev, frontend-dev, test-writer |
| bug, investigate, debug | Debug | 3-5 investigators with different hypotheses |
| review, code review | Review | security-reviewer, perf-reviewer, quality-reviewer |
| Other | Custom | Describe the structure |
Patterns detailed in .claude/skills/agent-teams/patterns.md.
Expected output
- Team: Agents created with roles and scopes
- Consolidated report: Results per agent + synthesis + priorities
- Cleanup: Confirmation of shutdown of all agents
Related agents
| Agent | Usage |
|---|---|
/work:work-explore | Explore BEFORE launching a team |
/work:work-plan | Plan BEFORE a Feature pattern |
/qa:qa-audit | Single-agent alternative to the audit |
IMPORTANT: Always check that Agent Teams is enabled before creating a team.
YOU MUST choose the pattern appropriate to the task.
YOU MUST clean up the team after use (shutdown + cleanup).
NEVER launch more than 5 agents without warning about the token cost.
NEVER have 2 agents work on the same file.
Think hard about the task decomposition before spawning the agents.