Running two or three AI coding agents at the same time sounds efficient until they start overwriting each other's files. Batty is a new open-source tool built to fix exactly that problem.
The project, built by a developer who uses Claude Code and Codex daily, adds a supervisor layer above your AI coding agents. You define a team structure in a YAML config file - an architect that plans work, a manager that dispatches tasks, and engineers that execute. Batty launches each role in a separate tmux pane (a terminal tool that runs multiple isolated shell sessions side by side) and adds test gating between steps: an agent's work doesn't move forward until the code actually compiles and tests pass.
The problem is real. When multiple agents share a repo without coordination, they collide on the same files, produce code that doesn't compile, and create more cleanup work than the task was worth in the first place. Batty's approach is structure rather than hope - each engineer agent gets its own isolated working area, and nothing advances without a quality check.
The YAML team definition is the notable design choice here. Low setup friction - describe the roles and their relationships once, and the tool handles orchestration from there. No per-project scripts to maintain.
This is early-stage open-source, not a polished product. The repo is new and built primarily around the author's own workflow with Claude Code and Codex. But the underlying pattern - hierarchical agent teams with quality gates between steps - is where the bigger players like Cursor are also heading. A solo developer shipping a working version of this for free is a good indicator that parallel agent coordination is becoming a common enough pain point to warrant dedicated tooling.
Available at github.com/battysh/batty.