Agent Teams Skill
Load with: base.md + code-review.md + security.md + session-management.md
Purpose: Every project initialized with claude-bootstrap runs as a coordinated team of AI agents. This is the default workflow, not optional. Teams enforce a strict TDD pipeline where no step can be skipped.
Requires: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 environment variable.
Core Principle
Every feature follows an immutable pipeline enforced by task dependencies:
┌─────────────────────────────────────────────────────────────────┐
│ STRICT FEATURE PIPELINE (IMMUTABLE) │
│ ────────────────────────────────────────────────────────────── │
│ │
│ 1. SPEC Write feature specification │
│ ↓ (Feature Agent) │
│ 2. REVIEW Quality Agent reviews spec completeness │
│ ↓ (Quality Agent) │
│ 3. TESTS Write failing tests for all acceptance criteria │
│ ↓ (Feature Agent) │
│ 4. RED VERIFY Quality Agent confirms ALL tests FAIL │
│ ↓ (Quality Agent) │
│ 5. IMPLEMENT Write minimum code to pass tests │
│ ↓ (Feature Agent) │
│ 6. GREEN VERIFY Quality Agent confirms ALL tests PASS + coverage│
│ ↓ (Quality Agent) │
│ 7. VALIDATE Lint + type check + full test suite │
│ ↓ (Feature Agent) │
│ 8. CODE REVIEW Multi-engine review, block on Critical/High │
│ ↓ (Code Review Agent) │
│ 9. SECURITY OWASP scan, secrets detection, dependency audit │
│ ↓ (Security Agent) │
│ 10. BRANCH+PR Create feature branch, stage files, create PR │
│ (Merger Agent) │
│ │
│ No step can be skipped. Task dependencies enforce ordering. │
│ Quality Agent verifies RED/GREEN transitions. │
│ Code Review + Security Agents gate the merge path. │
│ Merger Agent handles branching and PR creation. │
└─────────────────────────────────────────────────────────────────┘
Default Agent Roster
Every project spawns 5 permanent agents + N feature agents:
┌─────────────────────────────────────────────────────────────────┐
│ DEFAULT TEAM ROSTER │
│ ────────────────────────────────────────────────────────────── │
│ │
│ PERMANENT AGENTS (always present) │
│ ───────────────────────────────── │
│ Team Lead Orchestration, task breakdown, assignment │
│ Uses delegate mode - NEVER writes code │
│ │
│ Quality Agent TDD verification (RED/GREEN phases) │
│ Coverage gates (>= 80%) │
│ Spec completeness review │
│ │
│ Security Agent OWASP scanning, secrets detection │
│ Dependency audit, .env validation │
│ Blocks on Critical/High │
│ │
│ Code Review Agent Multi-engine code review │
│ Claude / Codex / Gemini / All │
│ Blocks on Critical/High │
│ │
│ Merger Agent Creates feature branches │
│ Stages feature-specific files only │
│ Creates PRs via gh CLI │
│ NEVER merges - only creates PRs │
│ │
│ DYNAMIC AGENTS (one per feature) │
│ ──────────────────────────────── │
│ Feature Agent Implements one feature end-to-end │
│ (x N features) Follows strict pipeline above │
│ Uses Ralph loops for implementation │
│ │
└─────────────────────────────────────────────────────────────────┘
| Agent | Role | Plan Mode | Can Edit Code |
|---|---|---|---|
| team-lead | Orchestration, task breakdown, assignment | No (delegate mode) | No |
| quality-agent | TDD verification, coverage gates | Yes | No (read-only) |
| security-agent | OWASP scanning, secrets detection | Yes | No (read-only) |
| review-agent | Multi-engine code review | Yes | No (read-only) |
| merger-agent | Branch creation, PR management | No | No (git only) |
| feature-{name} | Feature implementation (one per feature) | No | Yes |
Team Lead Responsibilities
The Team Lead is the orchestrator. It NEVER writes code.
- Read
_project_specs/features/*.mdto identify all features - Break each feature into the 10-task dependency chain (see below)
- Spawn one feature agent per feature
- Assign initial tasks (spec-writing) to feature agents
- Monitor TaskList continuously for progress and blockers
- Handle blocked tasks and reassignment
- Coordinate cross-feature dependencies
- Send
shutdown_requestto all agents when all PRs are created - Clean up the team when done
Delegate mode is mandatory. The team lead uses only:
- TeamCreate, TaskCreate, TaskUpdate, TaskList, TaskGet
- SendMessage (message, broadcast, shutdown_request)
- Read, Glob, Grep (for monitoring)
Feature Agent Workflow (MANDATORY)
Each feature agent MUST follow this exact sequence. Task dependencies enforce ordering - a feature agent cannot start step N+1 until step N is marked complete and verified.
Step 1: Write Spec
- Create
_project_specs/features/{feature-name}.md - Include: description, acceptance criteria, test cases table, dependencies
- Follow the atomic TODO format from base.md skill
- Mark task complete -> Quality Agent reviews
Step 2: Write Tests (RED Phase)
- Write test files based on spec's test cases table
- Tests MUST cover ALL acceptance criteria
- Import modules that don't exist yet (they will fail)
- Mark task complete -> Quality Agent verifies tests EXIST and FAIL
Step 3: Wait for RED Verification
- Quality Agent runs tests and verifies ALL new tests fail
- If any test passes without implementation -> rewrite tests
- Quality Agent marks verification complete -> unlocks implementation
Step 4: Implement (GREEN Phase)
- Write minimum code to make all tests pass
- Follow simplicity rules from base.md (20 lines/function, 200 lines/file, 3 params)
- Use Ralph loops (
/ralph-loop) for iterative implementation - Run tests after implementation - ALL must pass
- Mark task complete -> Quality Agent verifies tests pass
Step 5: Wait for GREEN Verification
- Quality Agent runs full test suite and checks coverage
- Coverage must be >= 80%
- If tests fail or coverage insufficient -> fix and re-request
- Quality Agent marks verification complete -> unlocks validation
Step 6: Validate
- Run linter (ESLint / Ruff)
- Run type checker (TypeScript / mypy)
- Run full test suite with coverage
- Fix any issues
- Mark task complete -> unlocks code review
Step 7: Wait for Code Review
- Code Review Agent runs
/code-reviewon changed files - If Critical or High issues -> fix and re-request review
- Code Review Agent marks complete -> unlocks security scan
Step 8: Wait for Security Scan
- Security Agent runs security checks
- If Critical or High issues -> fix and re-request scan
- Security Agent marks complete -> unlocks merge
Step 9: Wait for Branch + PR
- Merger Agent creates feature branch, stages files, creates PR
- Feature is complete when PR is created
Task Dependency Chain Model
For each feature "X", the team lead creates these 10 tasks with strict ordering:
┌────────────────────────────────────────────────────────────────┐
│ TASK CHAIN FOR FEATURE "X" │
│ │
│ Task 1: X-spec │
│ owner: feature-X │
│ blockedBy: (none) │
│ ↓ │
│ Task 2: X-spec-review │
│ owner: quality-agent │
│ blockedBy: X-spec │
│ ↓ │
│ Task 3: X-tests │
│ owner: feature-X │
│ blockedBy: X-spec-review │
│ ↓ │
│ Task 4: X-tests-fail-verify │
│ owner: quality-agent │
│ blockedBy: X-tests │
│ ↓ │
│ Task 5: X-implement │
│ owner: feature-X │
│ blockedBy: X-tests-fail-verify │
│ ↓ │
│ Task 6: X-tests-pass-verify │
│ owner: quality-agent │
│ blockedBy: X-implement │
│ ↓ │
│ Task 7: X-validate │
│ owner: feature-X │
│ blockedBy: X-tests-pass-verify │
│ ↓ │
│ Task 8: X-code-review │
│ owner: review-agent │
│ blockedBy: X-validate │
│ ↓ │
│ Task 9: X-security-scan │
│ owner: security-agent │
│ blockedBy: X-code-review │
│ ↓ │
│ Task 10: X-branch-pr │
│ owner: merger-agent │
│ blockedBy: X-security-scan │
└────────────────────────────────────────────────────────────────┘
Parallel Feature Execution
Multiple features run their chains in parallel. Shared agents process tasks as they unblock:
Feature: auth Feature: dashboard Feature: payments
auth-spec dash-spec pay-spec
auth-spec-review dash-spec-review pay-spec-review
auth-tests dash-tests pay-tests
auth-fail-verify dash-fail-verify pay-fail-verify
auth-implement dash-implement pay-implement
auth-pass-verify dash-pass-verify pay-pass-verify
auth-validate dash-validate pay-validate
auth-code-review dash-code-review pay-code-review
auth-security dash-security pay-security
auth-branch-pr dash-branch-pr pay-branch-pr
| | |
v v v
[All chains run simultaneously]
[Quality Agent handles all verify tasks as they unblock]
[Review Agent handles all review tasks as they unblock]
[Security Agent handles all scan tasks as they unblock]
[Merger Agent handles all branch-pr tasks as they unblock]
Inter-Agent Communication
Direct Messages (for targeted work)
Feature Agent -> Quality Agent: "Tests written for auth, ready for RED verify"
Quality Agent -> Feature Agent: "All 7 tests fail as expected. Proceed to implement"
Feature Agent -> Review Agent: "Implementation complete, ready for code review"
Review Agent -> Feature Agent: "2 High issues found: [details]. Fix before proceeding"
Security Agent -> Merger Agent: "Security scan passed for auth feature"
Merger Agent -> Team Lead: "PR #42 created for auth feature"
Task List (source of truth for state)
- All agents check TaskList after completing work
- Quality Agent claims verification tasks automatically
- Review Agent claims code-review tasks automatically
- Security Agent claims security-scan tasks automatically
- Merger Agent claims branch-pr tasks automatically
Broadcast (rare - blocking issues only)
- Team Lead -> All: "Blocking dependency found between auth and dashboard"
- Security Agent -> All: "Critical vulnerability in shared dependency"
Feature Agent Spawning
The team lead spawns one feature agent per feature:
- Read
_project_specs/features/*.md - For each feature spec, spawn a feature agent:
- name:
feature-{feature-name} - Uses
.claude/agents/feature.mddefinition - Spawn prompt includes the feature name and spec location
- name:
- Create the full 10-task dependency chain for that feature
- Assign the spec-writing task to the feature agent
Example
If project has 3 features: auth, dashboard, payments
- Spawn:
feature-auth,feature-dashboard,feature-payments - Create 30 tasks total (10 per feature)
- Each feature agent starts with their spec task
- All 3 work in parallel
Branch and PR Strategy
One branch per feature. One PR per feature.
Branch naming: feature/{feature-name}
PR title: feat({feature-name}): {short description}
PR body: Generated from spec + test results + review + security results
The Merger Agent:
git checkout main && git pull origin maingit checkout -b feature/{feature-name}- Stages ONLY files changed for this feature (never
git add -A) - Commits with descriptive message including verification results
git push -u origin feature/{feature-name}gh pr createwith full template including:- Summary from feature spec
- Test results from quality verification
- Code review summary from review agent
- Security scan results from security agent
- Checklist of all pipeline steps completed
Quality Gates
Workflow Enforcement (via task dependencies)
- Task dependencies make it structurally impossible to skip steps
- A feature agent cannot see "implement" until quality agent completes "tests-fail-verify"
- This is the primary enforcement mechanism
Cross-Agent Verification (trust but verify)
- Quality agent independently runs tests (doesn't trust feature agent's report)
- Security agent independently scans (doesn't trust review agent)
- Merger agent verifies all predecessor tasks are complete before branching
Blocking Rules
- Quality Agent: blocks if tests don't fail (RED) or don't pass (GREEN) or coverage < 80%
- Code Review Agent: blocks on Critical or High severity issues
- Security Agent: blocks on Critical or High severity findings
- Merger Agent: refuses to branch if any predecessor task is incomplete
Integration with Existing Skills
| Existing Skill | How Agent Teams Uses It |
|---|---|
| base.md | TDD workflow, atomic todos, simplicity rules - all agents follow |
| code-review.md | Review Agent executes /code-review per this skill |
| security.md | Security Agent follows OWASP patterns from this skill |
| session-management.md | Each agent maintains its own session state |
| iterative-development.md | Feature agents use Ralph loops for implementation |
| project-tooling.md | Merger Agent uses gh CLI for branches and PRs |
| team-coordination.md | Superseded by agent-teams for automated coordination |
Environment Setup
Required Setting
// settings.json or environment
{
"env": {
"CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
}
}
Project Structure (created by /initialize-project)
.claude/
agents/ # Agent definitions (from agent-teams skill)
team-lead.md
quality.md
security.md
code-review.md
merger.md
feature.md
skills/
agent-teams/ # This skill
SKILL.md
agents/ # Agent definition templates
base/
code-review/
security/
...
Spawning the Team
Automatic (via /initialize-project)
After project setup completes, Phase 6 asks for features and spawns the team automatically.
Manual (via /spawn-team)
For existing projects: run /spawn-team to spawn the team from existing feature specs.
Limitations
- Experimental feature - Agent teams require the experimental env var
- No nested teams - Teammates cannot spawn sub-teams
- One team per session - Clean up before starting a new team
- No session resumption - If session dies, re-run
/spawn-team(tasks persist) - File conflicts - Features sharing files must be serialized by team lead
- Token cost - Each agent is a separate Claude instance (5 + N instances)