You already use a coding agent. Maybe Claude Code, maybe Codex, maybe Cursor. One agent on one task works fine. The trouble starts when you run several at once: branches collide, CI breaks, and every fix waits on you.
Agent Orchestrator fixes that. It is a free, open source desktop app (Apache 2.0) that gives your agents a manager. You describe the outcome, it plans the work, spawns workers on their own branches, and sends failed CI back to the agent that owns the branch.
We put this kit together so you can go from zero to your first supervised agent team in one sitting. Let's go.
The Agent Orchestrator board. Source: the project's README.
Good to know before you start
It is still before version 1.0, so expect fast updates and the occasional rough edge.
It does not include any AI model. You bring your own agent subscriptions or API keys (Claude Code, Codex, Cursor and so on).
Code and execution stay on your machine. Cloud agents are on a waitlist, not generally available yet.
1. Install the desktop app
Grab the installer for your computer from the project's latest release on GitHub. There are builds for macOS (Apple silicon and Intel), Windows and Linux (AppImage, .deb and .rpm). The app checks for updates on its own.
On a Mac with Homebrew you can also run this one line in Terminal:
brew install agentwrapper/tap/agent-orchestratorMac tip: if macOS says the developer cannot be verified, right click the app, choose Open, then Open again.
You do not need the command line after this. The desktop app runs everything in the background for you.
2. Connect your agents
Agent Orchestrator does not replace your coding agent. It runs the ones you already have installed and signed in.
Make sure at least one agent works on its own first. For example, open a terminal and run
claudeorcodexand check it starts and is signed in.Open Agent Orchestrator and add a project. Point it at the Git repository you want your agents to work on.
When you create a task, pick the agent and model you want. You can mix them: Claude Code on one task, Codex on another.
It supports Claude Code, Codex, Cursor, OpenCode, Aider, GitHub Copilot, Gemini CLI, Goose and more than twenty others.
Creating a task: pick the agent and the model. Source: the project's README.
3. Your first orchestrator prompt
There are two ways to give work. New task sends one clear job to one worker. The orchestrator is a planning agent for the whole project: you describe an outcome, it breaks it into tasks, spawns workers (each on its own branch and worktree) and follows their progress.
Start small so you can watch how it works. Copy this into the orchestrator and adjust the brackets:
I want to [add a dark mode toggle to the settings page].
Before writing code, look at the repo and propose a plan split into
2 or 3 independent tasks that will not touch the same files.
Wait for my OK, then spawn one worker per task, each on its own branch.
Each worker should open a pull request and fix its own CI failures.Two more you can keep in your back pocket:
Review the open pull requests from today's workers. List anything
that conflicts with another branch and tell me which order to merge.Look at this repo and suggest the 3 highest impact improvements we
could ship this week. Explain the tradeoffs in plain language.
No code yet.
The orchestrator splitting one outcome into five workers. Source: the project's README.
4. Read the Kanban board
Every worker shows up as a card on one live board, and the card moves on its own based on what is really happening with its session, pull request, CI and reviews:
Working: the agent is busy, or ready for another instruction.
Needs you: something is blocked. Missing input, failed CI, requested changes. This is the only column you really have to watch.
In review: the pull request is open and waiting on checks or a review.
Ready to merge: approved work. Merge it and move on.
When CI fails or a reviewer asks for changes, Agent Orchestrator sends that feedback straight back to the agent that owns the branch, so you are not copying error logs between windows.
That's it. You now run a team, not a terminal.
The AI Leverage, by bestapps.ai on instagram
