Claude Code Agent Teams

Coordinate multiple Claude Code agents to work in parallel on independent tasks like multi-lens reviews, debugging, and full-stack features with defined file...

Security Scan

The skill appears to do what it claims—local orchestration helpers and templates for Claude Code agent teams—and contains only local helper scripts and documentation, but there are small metadata inconsistencies you should verify before installing.

Purpose & Capability

The skill's stated purpose (orchestrating Claude Code agent teams) matches the included templates and helper scripts. However, registry-level metadata you provided lists no required env vars/binaries while both SKILL.md and package.json declare a dependency on the 'claude' CLI and the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS env flag. This is a packaging/metadata inconsistency that should be clarified.

Instruction Scope

Runtime instructions are limited to enabling an experimental env flag, running the 'claude' CLI in PTY mode, pasting spawn prompts, and using local monitor/cleanup scripts. The scripts read local process state (pgrep/pkill), check PATH for binaries, and interact with local files under the skill workspace—no instructions to read arbitrary system credentials or post data to external endpoints are present.

Install Mechanism

There is no automated install spec; the package is instruction- and script-based only. That minimizes risk because nothing is downloaded or executed automatically. The README suggests installing via 'clawhub install', which you should inspect separately if you plan to use ClawHub.

Credentials

The skill only requires a boolean experimental feature flag (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS) and the 'claude' CLI—appropriate for its purpose. But the top-level registry metadata you provided lists no required env vars or binaries while SKILL.md and package.json declare them; this mismatch should be corrected so users know what privileges to expect.

Persistence & Privilege

The skill does not request permanent 'always' inclusion, does not modify other skills' configs, and only suggests updating a user settings file (~/.claude/settings.json) to set the experimental flag (a user-level change). The scripts perform graceful or forceful termination of local 'claude' processes, which is reasonable for cleanup utilities.

Assessment

This skill is coherent with its stated purpose and contains only local helper scripts and templates, but do the following before installing/using it: (1) Confirm the required preconditions—install the 'claude' CLI and set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 if you want teams; (2) Inspect the scripts (validate-setup.sh, team-monitor.sh, team-cleanup.sh) yourself—they use pgrep/pkill and interact with local terminals but do not contact external servers; (3) Be cautious if using the suggested 'clawhub install'—review what ClawHub would run during installation; (4) Ask the publisher to correct the registry metadata mismatch (registry says no required env/bins while SKILL.md/package.json declare them). If you need the skill to run in a restricted environment, consider running it in an isolated shell or container first to validate behavior.

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

Template	Use Case
parallel-review.md	Multi-lens code review (security + performance + tests)
competing-hypotheses.md	Debug by testing competing theories
fullstack-feature.md	Frontend/backend/tests coordination
architecture-decision.md	ADR with adversarial debate
bottleneck-analysis.md	Cross-domain performance investigation
inventory-classification.md	Data-parallel classification/refactoring

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

Approach	Best For
Single session	Sequential work, exploration
Agent teams (this skill)	Parallel independent tasks with coordination
OpenClaw subagents	Background work, long-running tasks

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