Agent Teams Advanced

Agent Teams Advanced Patterns

Claude Code Agent Teams enable parallel multi-agent coordination within a shared task context. This skill covers production-grade patterns for designing, scaling, and debugging agent team topologies.

Agent Teams Mechanics

Enablement:

• Set environment variable: CLAUDE_ENABLE_TEAMS=true
• Or pass --enable-teams flag to Claude Code CLI
• Each teammate gets an independent context window and git worktree

How teams differ from subagents:

• Subagents are sequential, fire-and-forget workers. Main agent waits for results then continues.
• Teammates run in parallel with peer-to-peer messaging. All teammates share a task list and can see each other's progress in real-time.
• Independent sessions: Each teammate maintains its own conversation state and can recover if it fails without blocking other teammates.
• Shared responsibility: Teammates collectively own the goal, not individual subtasks assigned by a parent.

Team Topology Patterns

Hub-and-Spoke (Lead-Driven)

One lead assigns work to teammates and collects results.

• Best for: Hierarchical workflows, clear task decomposition, reporting requirements
• Lead responsibilities: Break work into tasks, assign to teammates, merge results
• Teammates: Execute assigned tasks, report status and blockers
• Cost: 1 lead (Opus) + N teammates (Sonnet/Haiku)

      ┌─────────┐
      │  Lead   │
      └────┬────┘
      ┌────┴────┬─────────┬─────────┐
      │    │    │    │    │    │    │
      ▼    ▼    ▼    ▼    ▼    ▼    ▼
   Team1 Team2 Team3 Team4 Team5 Team6

Mesh Network (Peer-to-Peer)

All teammates can message each other directly. Useful for truly parallel, interdependent work.

• Best for: Feature branches, cross-domain coordination, research teams
• Communication: Direct peer messaging, no bottleneck at lead
• Self-organization: Teammates claim tasks from shared queue
• Cost: N agents (all Sonnet for parity)

   ┌──────────────────────────┐
   │                          │
   ▼                          ▼
Team1 ◄──────────────────► Team2
   │                          │
   │ ◄──────────────────►     │
   │                          │
   └──────────►Team3◄─────────┘
      ◄────────┐ │ ┌───────────►
               │ │ │
            Self-Claim Queue

Pipeline (Sequential Handoff)

Work moves through stages. Each stage completes before passing to the next.

• Best for: Transformations, migrations, multi-phase analysis
• Stages: Analyzer → Implementer → Validator → Deployer
• Blocking: Each stage waits for previous to complete
• Cost: 4+ agents (Opus-Sonnet-Sonnet-Sonnet typically)

Hierarchy (Multi-Level Teams)

Team leads manage sub-teams. Useful for large organizations.

• Best for: Enterprise projects, scaled parallel work
• Structure: Main team + 3-4 sub-teams, each with lead
• Complexity: Coordination overhead increases at each level
• Cost: Scales quadratically (main lead + leads × members)

Cross-Team Communication

Shared Task List

All teammates can view, claim, and complete tasks from a shared queue.

• Self-claiming: Teammates ask "who's claiming this?" and pick up the next unclaimed task
• Status visibility: All teammates see what everyone else is working on
• Prevents duplicate work: Task locks prevent two teammates from claiming the same work

Peer-to-Peer Messaging

Direct messages between teammates for synchronous coordination.

Teammate1 → "Finished schema, Team2 can now generate tests"
Teammate2 → "Tests running, Team3 estimate for performance review?"
Teammate3 → "Ready when schema is final—what's your ETA?"

Status Broadcasting

Periodic updates visible to entire team (every 5-10 minutes).

• Progress: "Completed 3/8 unit tests"
• Blockers: "Waiting on Team4's API changes"
• Handoff signals: "Feature branch ready for review"

Conflict Resolution

When two teammates claim the same task:

1. Task lock: First to claim gets exclusive lock (60 sec default)
2. Conflict detection: System alerts second claimer
3. Graceful degradation: Second teammate picks next task or supports first with parallel work

Git Worktree Coordination

Each teammate gets an isolated git worktree, preventing merge conflicts during work.

Worktree naming:

# Automatic naming by team ID and teammate index
.git/worktrees/team-{team_id}-{teammate_index}/

Merge strategy on completion:

1. Squash and merge: Default. Reduces commit graph noise
2. Rebase and merge: Preserves linear history, good for pipelines
3. Three-way merge: Default for independent branches, may have conflicts

Conflict handling:

• Auto-resolvable: Same file, different sections → merge succeeds
• Conflict markers: Both teammates edit same section → manual resolution required
• Fallback: Team lead (or designated resolver) reviews and resolves conflicts

Branch naming conventions:

feature/{team-id}/{teammate-role}
  examples: feature/cce-001/frontend, feature/cce-001/backend

bugfix/{team-id}/{issue-number}
  examples: bugfix/cce-002/123, bugfix/cce-002/456

Team Sizing Guidance

Cost scaling:

• Each teammate is a separate Opus/Sonnet session
• 2 Sonnet agents = ~2x cost, but work completes in ~50-60% of sequential time (depends on parallelization efficiency)
• 5 agents = 5x cost, but work may complete in ~30-40% of sequential time (coordination tax increases)

Cost-benefit formula:

TeamCost = (team_size × hourly_rate_per_agent)
Speedup = 100% / (1 + coordination_overhead)
ROI = (sequential_time - (team_time × speedup)) / TeamCost
→ ROI > 0 when team_time * speedup < sequential_time

Failure Handling

Single Teammate Failure

• Detection: No heartbeat for 30 seconds, or task marked as failed
• Impact: Other teammates continue. Orphaned work reassigned to next available teammate.
• Recovery: Failed teammate wakes up in a fresh session, picks next available task
• Max retries: 2 per task by default. After 2 failures, escalate to team lead.

Cascading Failure

When teammate A's failure blocks teammate B (dependency).

• Timeout: After 60 seconds of blockage, attempt workaround or parallel path
• Escalation: Involve team lead to decide: retry, skip, or reassign work
• Partial success: Mark work as "completed with warnings," document blockers

Team-Wide Failure

All teammates down or unresponsive.

• Fallback: Team lead takes over work sequentially
• Graceful degradation: Resume from last successful checkpoint (git commit)
• Monitoring: Alert user after 2 minutes of team inactivity

Monitoring Teammate Health

# Check teammate status
claude-code teams status --team-id cce-001

# Output:
# ┌─────────────┬────────────┬──────┬─────────────┐
# │ Teammate    │ Status     │ Task │ Last Update │
# ├─────────────┼────────────┼──────┼─────────────┤
# │ Frontend    │ RUNNING    │ 5/12 │ 2m ago      │
# │ Backend     │ COMPLETED  │ 8/8  │ 5m ago      │
# │ Tests       │ STALLED    │ 2/6  │ 10m ago     │
# └─────────────┴────────────┴──────┴─────────────┘

Custom Team Templates

Define team compositions beyond built-in templates using a team configuration file.

Example: Custom 3-person migration team

version: 1
name: "Database Migration"
description: "Schema analyzer, data migrator, validator"
team:
  - name: "analyzer"
    role: "Database schema analysis"
    model: "claude-sonnet-4-6"
    skills:
      - sql-analysis
      - schema-design
    timeout_seconds: 1800

  - name: "migrator"
    role: "Data migration execution"
    model: "claude-sonnet-4-6"
    skills:
      - data-migration
      - transformation-engine
    timeout_seconds: 3600

  - name: "validator"
    role: "Post-migration verification"
    model: "claude-haiku-4-5"
    skills:
      - data-validation
      - integrity-checking
    timeout_seconds: 900

communication: "mesh"
task_strategy: "shared_queue"
parallelization: "full"

Real-World Patterns

1. Feature Branch Team (3 agents)

• Frontend (Sonnet): UI components, styling, state management
• Backend (Sonnet): API endpoints, database schema changes
• Tests (Haiku): Unit, integration, and E2E tests

→ Parallel development with synchronized handoffs

2. Migration Team (4 agents)

• Analyzer (Sonnet): Audit old codebase, create migration plan
• Migrator (Sonnet): Execute refactoring
• Validator (Haiku): Verify no regressions
• Lead (Opus): Coordinate, resolve conflicts, review PRs

→ High confidence, rapid migrations

3. Incident Response Team (3 agents)

• Triager (Haiku): Investigate error signals, classify severity
• Fixer (Sonnet): Implement hotfix
• Verifier (Haiku): Confirm fix resolves issue, no side effects

→ Fast incident turnaround

4. Research Team (4 agents + synthesizer)

• Researcher 1-3 (Haiku): Parallel investigation of 3 topics
• Synthesizer (Sonnet): Combines findings, creates coherent summary

→ Cover more ground in parallel research

5. Review Board (3 agents)

• Security Reviewer (Sonnet): Audit for vulnerabilities
• Performance Reviewer (Haiku): Check for bottlenecks, inefficient patterns
• Quality Reviewer (Haiku): Code style, maintainability, test coverage

→ Parallel review perspectives

Cost Management

Estimating Team Cost

Base cost = (num_teammates × model_cost_per_hour) × estimated_duration_hours
Example: 3 Sonnet agents × $15/hour × 2 hours = $90

vs.

Sequential cost = 1 Sonnet × $15/hour × 6 hours = $90
Speedup: 3 agents complete in 2 hours vs. 6 sequential → 3x faster, same cost

Model Assignment Per Teammate

• Opus (team lead): Complex architecture, conflict resolution, synthesis
• Sonnet (implementation): Main work — features, refactoring, migrations
• Haiku (research, validation): Lightweight tasks — checking, summarizing, testing edge cases

When Teams Are Worth the Cost

• YES, use teams if:

  • Work is truly parallelizable (frontend + backend, not cascading dependencies)
  • Task duration > 1 hour (coordination overhead justified)
  • Deadline is critical (2-3x faster worth the cost)
  • Risk is high (parallel review + validation worth extra cost)

• NO, use sequential if:

  • Tasks are cascading (B waits for A, C waits for B)
  • Work is < 30 minutes (coordination overhead too high)
  • Budget is fixed and limited (sequential is cheaper)
  • High variance in teammate performance (uneven completion)

Cost Optimization Tips

1. Use Haiku for validation instead of Sonnet — 75% cheaper, similar quality
2. Limit team size to 3-4 unless truly independent tracks exist
3. Batch small tasks instead of assigning individually (reduced coordination overhead)
4. Set aggressive timeouts (900 seconds) to fail fast on stuck teammates
5. Monitor actual speedup — if actual duration > 50% of sequential, reduce team size

See also: teams-architect agent for interactive team design assistance.

Source: markus41/claude — distributed by TomeVault.