agent-teams

This Skill addresses the lack of coordination mechanisms when multiple Claude processes work in parallel. In practical development, when multiple Claude agents need to handle different modules simultaneously (e.g., frontend/backend parallel development, multi-party code review, large-scale refactoring), the absence of a unified communication and task distribution method can easily lead to conflicts or duplicated efforts. Agent Teams implements a lightweight, network‑free collaboration layer using a pure filesystem protocol (JSON files + fcntl locks). Each agent runs independently in a tmux pane, exchanging messages and tasks through shared file queues.

Usage is straightforward: create a team configuration file (containing member list and roles), then launch a separate claude process for each team member with their assigned identity. The system automatically generates inboxes and tasks directories under ~/.claude/teams/<team-name>/. Agents poll their own inbox files to retrieve tasks, with atomic writes and file locks ensuring concurrency safety. You can then assign tasks, synchronize progress, or aggregate results using natural‑language instructions, such as “Have the frontend agent complete the login page, the backend agent implement the API endpoints, then summarize.”

It is particularly suitable for teams or individuals requiring parallel multi‑agent collaboration, especially those already using the Claude CLI for large‑scale projects. Typical scenarios include: cross‑module parallel development, multi‑perspective code review, and coordinating changes across multiple subsystems during complex refactoring. For developers already working in a tmux environment, a collaboration network can be set up quickly with zero network dependencies.

We recommend using this tool consistently during phases with heavy parallel tasks to effectively reduce context‑switching overhead. Note that it is primarily designed for loosely‑coupled parallel tasks and is not suitable for scenarios requiring tight sequential reasoning (e.g., step‑by‑step debugging). It also depends on tmux and the Claude CLI, so ensure those are installed. Additionally, the filesystem‑lock mechanism may become a performance bottleneck under high‑frequency concurrency; we suggest limiting the number of agents per team to 10 or fewer.