Meet Batty: The Orchestrator for Multi-Agent Coding Teams

Managing multiple AI coding agents simultaneously can be akin to herding cats. Each one stepping on the other's toes, leaving you more frazzled than productive. Enter Batty, a supervisor layer that's changing the game for developers juggling multiple AI assistants like Claude Code and Codex.

The Batty Blueprint

Batty isn't just another agent framework. It's the orchestrator that brings order to your chaotic multi-agent world. The magic starts with a simple YAML file. Define your team: an architect to plan, a manager to dispatch, and engineers to execute. Each role gets its own tmux pane, and engineers work in isolated git worktrees, reducing the risk of file conflicts.

Crucially, Batty isn't embedding any models. It thrives on orchestrating existing agent CLIs using tmux for the runtime and git worktrees for task isolation. The config's all YAML, the kanban board is Markdown, inboxes are Maildir, and logs are JSONL. Want to see your team's state? Just `git diff` it.

Built in Rust and published on crates.io (v0.1.0), Batty runs a synchronous 5-second poll loop. It watches pane output to spot idle, active, or dead agents, reads session files on disk to avoid false idle detection, and uses a merge lock for concurrent worktree merges. Simplicity over async complexity. Ship it to testnet first. Always.

Why It's a Game Changer

Running 3-5 engineers in parallel hits the sweet spot. Beyond that, the codebase becomes the bottleneck. More agents don't equal better results. It's about task decomposition quality. A solid architect prompt is more valuable than a swarm of agents tackling poorly-defined tasks.

Test gating is the secret sauce here. Without it, agents wrap up work that breaks everything downstream. With it, chaos gives way to coherent completion. Yet, don't think of Batty as a fire-and-forget solution. It's more like managing a junior team, where your job is to supervise multiple workstreams rather than doing one.

Some might wonder if this is overkill. But if you've ever felt the pain of coordinating AI agents, you'll see why Batty's approach is compelling. Tmux-IDE and vibe-kanban offer alternatives, but Batty's focus on supervision through test gating and communication constraints stands out. Different tools for different workflows.

What's Next for Batty?

It's early days for Batty, with its core loop solid but its API still evolving. The eight built-in templates range from a solo agent setup to a large-scale operation with 19 agents and multiple management layers. The architecture diagram in the README lays out the supervision flow in detail. Check out the 2-minute demo linked in the docs to see Batty in action.

For developers running multiple agents, what's your biggest operational pain point? Batty might just have the answer.