Agent Teams

Agent Teams Coordination

Overview

Claude Code Agent Teams enables multiple Claude CLI processes to collaborate on shared work through a filesystem-based coordination protocol. Each teammate runs as an independent claude process in a tmux pane, communicating via JSON files guarded by fcntl locks, with no database, daemon, or network layer.

This skill provides the patterns for orchestrating agent teams effectively.

When To Use

• Parallel implementation across multiple files or modules
• Multi-agent code review (one agent reviews, another implements fixes)
• Large refactoring requiring coordinated changes across subsystems
• Tasks with natural parallelism that benefit from concurrent agents

When NOT To Use

• Single-file changes or small tasks (overhead exceeds benefit)
• Tasks requiring tight sequential reasoning (agents coordinate loosely)
• When claude CLI is not available or tmux is not installed

Prerequisites

bashCopy code
# Verify Claude Code CLI
claude --version

# Verify tmux (required for split-pane mode)
tmux -V

# Enable experimental feature (set by spawner automatically)
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Protocol Architecture

Copy code
~/.claude/
  teams/<team-name>/
    config.json            # Team metadata + member roster
    inboxes/
      <agent-name>.json    # Per-agent message queue
      .lock                # fcntl exclusive lock
  tasks/<team-name>/
    1.json ... N.json      # Auto-incrementing task files
    .lock                  # fcntl exclusive lock

Design principles:

• Filesystem is the database: JSON files, atomic writes via tempfile + os.replace
• fcntl locking: Prevents concurrent read/write corruption on inboxes and tasks
• Numbered tasks: Auto-incrementing IDs with sequential file naming
• Loose coupling: Agents poll their own inbox; no push notifications

Quick Start

1. Create a Team

bashCopy code
# Programmatic team setup (via MCP or direct API)
# Team config written to ~/.claude/teams/<team-name>/config.json

The team config contains:

• name, description, created_at (ms timestamp)
• lead_agent_id, lead_session_id
• members[]: array of LeadMember and TeammateMember objects

2. Spawn Teammates

Each teammate is a separate claude CLI process launched with identity flags:

bashCopy code
claude --agent-id "backend@my-team" \
       --agent-name "backend" \
       --team-name "my-team" \
       --agent-color "#FF6B6B" \
       --parent-session-id "$SESSION_ID" \
       --agent-type "general-purpose" \
       --model sonnet

See modules/spawning-patterns.md for tmux pane management and color assignment.

3. Create Tasks with Dependencies

jsonCopy code
{
  "id": "1",
  "subject": "Implement API endpoints",
  "description": "Create REST endpoints for user management",
  "status": "pending",
  "owner": null,
  "blocks": ["3"],
  "blocked_by": [],
  "metadata": {}
}

See modules/task-coordination.md for state machine and dependency management.

4. Coordinate via Messages

jsonCopy code
{
  "from": "team-lead",
  "text": "API endpoints are ready for integration testing",
  "timestamp": "2026-02-07T22:00:00Z",
  "read": false,
  "summary": "API ready"
}

See modules/messaging-protocol.md for message types and inbox operations.

Coordination Workflow

1. agent-teams:team-created: Initialize team config and directories
2. agent-teams:teammates-spawned: Launch agents in tmux panes
3. agent-teams:tasks-assigned: Create tasks with dependencies, assign owners
4. agent-teams:coordination-active: Agents claim tasks, exchange messages, mark completion
5. agent-teams:team-shutdown: Graceful shutdown with approval protocol

Crew Roles

Each team member has a role that determines their capabilities and task compatibility. Five roles are defined: implementer (default), researcher, tester, reviewer, and architect. Roles constrain which risk tiers an agent can handle. See modules/crew-roles.md for the full capability matrix and role-risk compatibility table.

Team Formation

For mission-level team sizing, use the Team Formation rules from references/team-formation.md. This defines:

• Role definitions: Coordinator (mission lead), Agents (task owners), Reviewer (adversarial challenger)
• Team sizing rules: Simple (1), Moderate (2-4), Complex (5-7), Critical (5-10)
• Maximum team size: 10 agents (coordination overhead limit)
• File ownership rules: Prevent conflicts with clear ownership

See references/team-formation.md for full team sizing guidance and example team formations.

Health Monitoring

Team members can be monitored for health via heartbeat messages and claim expiry. The lead polls team health every 60s with a 2-stage stall detection protocol (health_check probe and 30s wait). Stalled agents have their tasks released and are restarted or replaced following the "replace don't wait" doctrine. See modules/health-monitoring.md for the full protocol and state machine.

Module Reference

• team-management.md: Team lifecycle, config format, member management
• messaging-protocol.md: Message types, inbox operations, locking patterns
• task-coordination.md: Task CRUD, state machine, dependency cycle detection
• spawning-patterns.md: tmux spawning, CLI flags, pane management
• crew-roles.md: Role taxonomy, capability matrix, role-risk compatibility
• health-monitoring.md: Heartbeat protocol, stall detection, automated recovery

Integration with Conjure

Agent Teams extends the conjure delegation model:

Use Skill(conjure:delegation-core) first to determine if the task benefits from multi-agent coordination vs. single-service delegation.

Delegation is on by default there, so the question a team formation answers is which shape of delegation fits, not whether to delegate. A teammate holding execution work delegates it the same way a single session would. A teammate holding design work keeps it, under the same Keep Local clauses.

When a teammate's delegation returns a fallback_reason, that teammate does the work itself and reports which providers it tried. It does not message the crew for a reassignment: no other teammate has a provider this one lacks, because the registry is machine-wide.

Worktree Isolation Alternative (Claude Code 2.1.49+)

For parallel agents that modify files, isolation: worktree provides a lightweight alternative to filesystem-based coordination. Each agent runs in its own temporary git worktree, eliminating the need for fcntl locking or inbox-based conflict avoidance on shared files.

• When to prefer worktrees over agent teams messaging: Agents work on overlapping files but don't need mid-execution communication
• When to prefer agent teams messaging: Agents need to coordinate discoveries or adjust plans based on each other's progress
• Combine both: Use agent teams for coordination with isolation: worktree per teammate for filesystem safety

Troubleshooting

Common Issues

tmux not found Install via package manager: brew install tmux / apt install tmux

Stale lock files If an agent crashes mid-operation, lock files may persist. Remove .lock files manually from ~/.claude/teams/<team>/inboxes/ or ~/.claude/tasks/<team>/

Orphaned tasks Tasks claimed by a crashed agent stay in_progress indefinitely. Use modules/health-monitoring.md for heartbeat-based stall detection and automatic task release. The health monitoring protocol detects unresponsive agents within 60s and 30s probe window and releases their tasks for reassignment.

Message ordering Filesystem timestamp resolution varies (HFS+ = 1s granularity). Use numbered filenames or UUID-sorted names to avoid collision on rapid message bursts.

Model errors on Bedrock/Vertex/Foundry (pre-2.1.39) Teammate agents could use incorrect model identifiers on enterprise providers, causing 400 errors. Upgrade to Claude Code 2.1.39+ for correct model ID qualification across all providers.

Nested session guard (2.1.39+) If claude refuses to launch within an existing session, ensure you're using tmux pane splitting (not subshell invocation). The guard is intentional. See modules/spawning-patterns.md for details.

Exit Criteria

• Team created with config and directories
• Teammates spawned and registered in config
• Tasks created with dependency graph (no cycles)
• Agents coordinating via inbox messages
• Graceful shutdown completed