agent-teams

This Skill addresses the lack of a unified coordination mechanism when multiple Claude processes collaborate. In practical development, when multiple AI agents are needed to handle different subtasks in parallel (e.g., frontend/backend separation, simultaneous code review and fixes), the absence of a lightweight coordination layer often forces developers to manually manage sessions and pass context, leading to confusion and conflicts.

Usage is straightforward: simply define a team configuration file (containing member list and roles), then start each companion process using parameters like claude --agent-id. The system will automatically create JSON files for inboxes and tasks in the ~/.claude/teams/ directory. It uses fcntl file locks to ensure concurrency safety, with each agent independently polling its own task queue, enabling loosely-coupled asynchronous collaboration.

It is particularly suitable for teams or individuals working on parallel multi-module development, large-scale refactoring, or multi-agent code reviews. This is especially useful for those already using the Claude CLI for coding assistance who want to organize multiple AI agents into a virtual team to handle cross-file changes, cross-module testing, or heterogeneous tasks (e.g., frontend + backend + testing).

We recommend using this tool when tasks have inherent parallelism and each subtask can be completed independently. Note that it is primarily designed for loosely-coupled collaboration scenarios and is not suitable for tasks requiring tight sequential reasoning (e.g., single-line code debugging). It also requires local installation of tmux and the Claude CLI.