Agency
A desktop app for running coding agents in parallel, each one isolated in its own git worktree. Watch them work in real terminals, review what they changed, merge the ones that got it right. Free, open source, and used to build itself.
getagency.dev ↗
The problem
Running one coding agent is easy. You open a terminal, point it at a repo, and watch. Running the second one is where it falls apart: two agents in the same working tree overwrite each other’s edits, fight over the branch, and leave you with a diff nobody can attribute. So most people run one at a time and wait, which wastes the one thing an agent is actually good for — being several places at once.
The workaround everybody reaches for is git worktrees, by hand: cut a worktree, open a terminal, launch the agent, remember which of the six windows is which, and merge them back one at a time. That works, and it stops working at about three.
We wanted the version where the branching, the terminals, the review and the merge are one window.
What we built
Agency runs the coding agent CLIs you already use, side by side, on your own machine. Every run opens in its own git worktree on its own branch, so agents working the same repo never touch each other’s files. Watch them all in a grid, or pull one full size when it needs you.
It does not reimplement your agent. Each run is a real PTY with the CLI’s own interface in it, so the TUI works, you can interrupt it, and you can type. Eleven agent CLIs are built in and detected on your PATH — Agency offers to install the ones that are missing — and your keys and subscriptions stay exactly as they were.
Around that sit the parts a parallel workflow needs. A local issue tracker that is a folder of markdown files in your project, so an agent can pick an issue up, move it, comment on it and file a follow-up, because all of those are just edits. A wiki in the same repo, with an agent available in the side panel of any note. Per-run diff review, PR creation and conflict resolution, so you read the change before you trust it. Desktop alerts when a run finishes, goes quiet waiting on you, or hits a conflict.
It runs on macOS 11 or later on Apple Silicon and on Linux, x86_64 and arm64. It is free, Apache-2.0, pre-1.0 and in beta, and there is no account. No analytics, no telemetry, no backend of ours: the only call Agency makes on its own is a version check against GitHub’s public API, and that is a toggle in Settings.
Built with itself
The first branch the app cut for itself merged into main nineteen days after the first commit, and from that point the project ran on itself. Eleven weeks from first commit to beta, 788 commits, 163 of them merges from agent/* branches cut by the app, against 167 issues in its own markdown tracker.
That is the part we care about as a studio. Dogfooding is not a virtue signal here — it is the only honest test of a tool whose whole job is keeping several agents useful at once. Every rough edge in dispatch, review and merge showed up in our own week before it could show up in yours, and the commit history is the receipt.
Why it matters for client work
Client projects are where the method gets paid for. Agency is how we run them: an issue queue in the repo, several agents working separate branches against it, and a human reading every diff before it merges.
That is what makes a fixed price honest. The speed comes from parallelism and from never having to wait on a handoff. The quality comes from the same review step it always did — one person, reading the change, refusing the ones that are wrong.
Agency is at getagency.dev, free to download.
If a tool cannot survive being used to build itself, it is not ready for anybody else's work either. Agency wrote most of Agency, on branches Agency cut, and we reviewed every one of them in the app.
