Team Coordination Skill

Purpose

Orchestrate Claude Code Agent Teams for SAFe multi-agent workflows. This skill provides patterns for spawning teams, assigning work by role, enforcing quality gates via task dependencies, and managing the full delivery pipeline.

Prerequisites

Agent Teams are experimental and must be enabled:

json
// .claude/settings.json or settings.local.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

When This Skill Applies

Invoke this skill when:

• Orchestrating work across multiple agent roles (FE, BE, QAS, etc.)
• Spawning a team for a Feature or Epic-level deliverable
• Coordinating parallel implementation with quality gate enforcement
• Running competing hypothesis debugging with multiple agents
• Performing parallel code review across security, performance, and test coverage

Agent Teams vs Subagents vs Background Agents

| Approach | Communication | Coordination | Best For | |----------|--------------|--------------|----------| | Agent Teams | DMs, broadcasts | Shared TaskList with dependencies | Complex multi-role SAFe workflows | | Subagents | Report back only | Main agent manages | Focused tasks, results only | | Background Agents | None | None | Fire-and-forget parallel work |

Use Agent Teams when: teammates need to share findings, challenge each other, and coordinate via SAFe gates. Use Subagents when: you need quick, focused workers that report back. Use Background Agents when: tasks are independent and don't need coordination.

SAFe Team Patterns

Pattern 1: TDM as Team Lead

The TDM (Technical Delivery Manager) is the natural team lead in the 11-agent model:

Create an agent team for {{TICKET_PREFIX}}-XXX implementation.

Team lead (TDM) responsibilities:
- Spawn teammates by role (BE, FE, QAS, etc.)
- Create tasks with SAFe gate dependencies
- Monitor progress and steer teammates
- Synthesize results and escalate blockers

Spawn these teammates:
- BE Developer: Implement API endpoints per spec
- FE Developer: Implement UI components per spec
- QAS: Validate acceptance criteria after implementation

Pattern 2: Task Dependencies for Quality Gates

Use addBlockedBy / addBlocks to enforce the SAFe pipeline:

Task structure for {{TICKET_PREFIX}}-XXX:

1. "Implement API endpoint" (owner: be-developer)
2. "Implement UI components" (owner: fe-developer)
3. "QAS validation" (owner: qas, blockedBy: [1, 2])
4. "Create PR" (owner: rte, blockedBy: [3])
5. "Stage 1 review" (owner: system-architect, blockedBy: [4])

This enforces: Implementation → QAS → RTE → Architect Review

Pattern 3: Spawning Teammates by Role

Map harness agent roles to teammate configurations:

Spawn teammates with role-specific prompts:

BE Developer teammate:
- Load: api-patterns, rls-patterns skills
- Task: Implement endpoints per spec at specs/{{TICKET_PREFIX}}-XXX-spec.md
- Constraint: All DB operations must use RLS context helpers

FE Developer teammate:
- Load: frontend-patterns skill
- Task: Implement UI per spec
- Constraint: Follow component patterns in patterns_library/ui/

QAS teammate:
- Load: testing-patterns skill
- Task: Execute testing strategy from spec
- Constraint: Must verify all acceptance criteria before approving
- Plan approval required (require plan approval before QAS makes changes)

Pattern 4: Parallel Code Review

Create an agent team to review PR #XXX. Spawn three reviewers:
- Security reviewer: Focus on RLS enforcement, input validation, auth checks
- Architecture reviewer: Focus on pattern compliance, separation of concerns
- Test reviewer: Focus on test coverage, edge cases, acceptance criteria

Have them each review independently, then share and challenge findings.

Pattern 5: Competing Hypothesis Debugging

Users report [issue description]. Spawn 3-4 teammates to investigate:
- Teammate 1: Investigate [hypothesis A]
- Teammate 2: Investigate [hypothesis B]
- Teammate 3: Investigate [hypothesis C]

Have them talk to each other to challenge and disprove theories.
Update findings as consensus emerges.

Communication Patterns

Direct Messages (Most Common)

Send a message to the BE developer:
"The API endpoint needs to handle pagination. Check patterns_library/api/
for the standard pagination pattern before implementing."

Broadcast (Use Sparingly -- Expensive)

Broadcast to all teammates:
"STOP: Architecture decision changed. The auth middleware now uses
{{AUTH_PROVIDER}} session tokens instead of JWT. Check the updated
spec before continuing."

Only broadcast for:

• Critical blocking issues
• Architecture changes affecting everyone
• Stop-the-line announcements

Shutdown Coordination

When all tasks are complete:
1. Send shutdown_request to each teammate
2. Wait for approval from each
3. Clean up team resources via TeamDelete

Quality Gate Hooks

TeammateIdle Hook

Validates teammates completed their work before going idle:

json
{
  "hooks": {
    "TeammateIdle": [{
      "command": "bash -c 'echo \"Verify all assigned tasks are completed before going idle\"'",
      "description": "Validate teammate completed assigned work"
    }]
  }
}

Exit code 2 sends feedback and keeps the teammate working.

TaskCompleted Hook

Validates task output meets criteria before allowing completion:

json
{
  "hooks": {
    "TaskCompleted": [{
      "command": "bash -c 'echo \"Verify acceptance criteria met before marking complete\"'",
      "description": "Validate task meets acceptance criteria"
    }]
  }
}

Exit code 2 prevents completion and sends feedback.

Team Sizing Guidelines

| Work Scope | Recommended Size | Tasks per Teammate | |-----------|-----------------|-------------------| | Single Story | 2-3 teammates | 3-4 tasks each | | Feature (multi-story) | 3-5 teammates | 5-6 tasks each | | Epic (parallel features) | 5-8 teammates | 5-6 tasks each |

Rules of thumb:

• Start with fewer teammates; scale up if needed
• Each teammate should own different files (avoid conflicts)
• 5-6 tasks per teammate keeps everyone productive
• More than 8 teammates rarely helps (coordination overhead)

Known Limitations

• Experimental: Enable via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS
• No session resumption: /resume doesn't restore in-process teammates
• One team per session: Clean up before starting a new team
• No nested teams: Teammates cannot spawn their own teams
• Permissions inherit: All teammates start with lead's permission mode
• Token cost: ~7x single session in plan mode

Authoritative References

• Official Docs: Agent Teams
• Hooks Reference: Hooks
• Subagents Comparison: Subagents
• Harness Agents: AGENTS.md
• Team Config: .claude/team-config.json