Claude Code Agent Teams

Orchestrate multiple Claude Code agents working together on complex tasks. Agents coordinate through a shared task list, message each other directly, and work in parallel on independent scopes.

When to Use

✅ Use for parallel independent work:

• Multi-lens code reviews (security, performance, tests in parallel)
• Debugging with competing hypotheses (test theories simultaneously)
• Full-stack features (frontend/backend/tests by separate specialists)
• Architecture decisions (adversarial debate produces stronger ADRs)
• Cross-domain investigations (bottleneck analysis across systems)
• Data-parallel work (inventory classification, bulk refactoring)

❌ Don't use for:

• Sequential work with tight dependencies
• Same-file edits (conflict risk)
• Simple tasks (coordination overhead > benefit)
• Learning/exploration (single agent with full context is better)

Rule of thumb: Teams multiply token cost 3-4x. Use only when parallelization benefit exceeds coordination overhead.

Quick Start

1. Enable Feature

Environment variable:

export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1

Or settings file (~/.claude/settings.json):

{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}

2. Validate Setup

./scripts/validate-setup.sh

Checks for: env var set, claude CLI available, tmux installed (optional).

3. Pick Template

Each template includes: when to use, team structure, copy-paste spawn prompt, expected outcomes.

4. Spawn Team

claude --pty
# Paste spawn prompt from template

Monitor with ./scripts/team-monitor.sh (separate terminal) or Ctrl+T (in Claude Code).

Core Principles

1. Clear file boundaries - Each teammate owns different files/directories (prevents conflicts)
2. Use delegate mode - Lead coordinates, teammates implement (include in spawn prompt)
3. Size tasks appropriately - 5-6 self-contained tasks per teammate
4. Wait for completion - Don't let lead merge results while teammates are mid-task
5. Start with research - Spawn researchers before builders to establish shared understanding

Cost Management

Agent teams multiply token usage 3-4x vs single session. Each teammate has its own context window.

Optimize with:

• Model mixing - Opus for lead (strategic), Sonnet for teammates (tactical) = ~40% savings
• Minimize broadcasts - Direct messages only (broadcasts copy to all teammates)
• Right-size teams - 3 teammates usually optimal; 6+ only for highly parallel work

See references/cost-management.md for detailed strategies.

Troubleshooting

Teammates not spawning?

• Verify CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set
• Restart Claude Code after setting env var

File conflicts?

• Define explicit file boundaries in spawn prompt (directory-level ownership)

Lead implementing instead of coordinating?

• Add "Use delegate mode: coordinate but don't implement" to spawn prompt
• Or press Shift+Tab to cycle into delegate mode

See references/troubleshooting.md for full list.

Automation

Monitor progress:

./scripts/team-monitor.sh

Cleanup (graceful shutdown):

./scripts/team-cleanup.sh

Display Modes

In-process (default) - All teammates in main terminal

• Shift+Up/Down - Select teammate
• Ctrl+T - Toggle task list
• Escape - Interrupt teammate

Split-pane (requires tmux or iTerm2) - Each teammate gets own pane

• Set teammateMode: "tmux" in ~/.claude/settings.json

References

• templates/ - 6 copy-paste ready spawn prompts with usage guidance
• scripts/ - Automation helpers (validate, monitor, cleanup)
• references/best-practices.md - When to use teams, task sizing, file boundaries
• references/cost-management.md - Token optimization strategies
• references/troubleshooting.md - Common issues and solutions
• examples/ - Real-world walkthroughs (PR review, bug hunt, feature build)

Comparison

Contributing

Found a useful pattern? Submit PR with:

1. Template or example
2. When to use + team structure
3. Spawn prompt + expected outcomes

Credits

Built with insights from Claude Code documentation, ClaudeFast patterns, and OpenClaw community.

Repository: https://github.com/matthew-a-gordon/claude-code-teams
License: MIT