agent-teams

This Skill addresses the pain point where multiple Claude CLI processes cannot efficiently collaborate. When developing large projects or handling multiple modules in parallel, a single Claude session is limited by serial execution, while team collaboration lacks a lightweight communication mechanism. Agent Teams enables each Claude agent to operate as an independent team member via a filesystem protocol, using JSON files to exchange messages, tasks, and states—eliminating the need for databases, daemons, or network layers, thereby significantly reducing the complexity of multi-agent collaboration.

Usage is straightforward: first, create a team configuration file (containing metadata such as name and member list) via the API or MCP. Then, use the standard claude command to specify identity parameters for each agent (e.g., --agent-id backend@my-team --team-name my-team). The system automatically launches independent Claude processes in tmux panes. Agents communicate via JSON files under ~/.claude/teams/<team-name>/, with fcntl locks ensuring concurrency safety. You can then orchestrate tasks using natural language instructions, such as “Have the backend agent implement a user login API while the frontend agent writes the corresponding React component.” Both agents will work in parallel and eventually exchange results via files.

It is well-suited for teams and individuals requiring parallel development, multi-agent code reviews, or large-scale refactoring—especially developers already using Claude CLI and familiar with tmux, as well as agent developers and project managers who need to coordinate multiple AI agents for complex projects (e.g., synchronous frontend/backend development, cross-module modifications). For teams already relying on Claude Code as their primary coding assistant, Agent Teams integrates naturally into existing workflows, upgrading solo efforts into coordinated cluster operations.

Recommended for use within tmux on Linux/macOS environments, ensuring the Claude CLI version supports parameters like --agent-id. This approach is primarily suitable for loosely coupled tasks—if tasks require tight sequential reasoning (e.g., recursive algorithms), the loose file-based communication between agents may introduce context-switching overhead. Additionally, filesystem I/O latency makes it unsuitable for high-frequency, real-time interaction scenarios, though it is fully adequate for most parallel development tasks.