How to run several AI agents on one project
Running several agents on one project needs three things: a machine to host them, a directory they share, and a rule deciding who writes where. Without the third, the second agent erases the first with nothing to announce it.
Step by step
- Create the project. A canvas starts its own machine and a private repository. That is where the shared directory lives, and it is the same one for every agent you open next.
- Connect your accounts. Sign in with the subscription or API key you already pay for at Anthropic, OpenAI or elsewhere. No tokens are resold, and usage stays on your account.
- Open a window per agent. Claude Code, Codex, Kimi and OpenCode each open in their own window with a real terminal. Two windows on Solo, five on Pro.
- Split by file, not by feature. Give each agent a corner of the project rather than a task that crosses all of it. Two agents on two areas move in parallel; two agents on one file wait for each other.
- Let the board arbitrate. An agent claims the paths it is about to touch before it writes, and the server refuses everybody else until it lets go. You can see who holds what, and who is waiting.
What you need first
A repository, even an empty one, and an account at a model provider. The rest comes with the project: the machine, the directory, and the board that hands out files.
An instructions file at the root of the repository changes the result more than any setting. Agents read it on startup, and it saves each of them the same first five mistakes.
The split that works
The useful unit is the file, not the feature. One agent on the API, one on the interface, one on the tests: they rarely cross and almost never wait. Three agents on "the checkout page" will fight over the same three files for an hour.
What you watch while it runs
Terminals scroll in their windows, and the board shows who holds which path. A blocked agent is visible before it does damage, which is the real difference from an invisible queue. The mechanism in full.
Straight answers
How many agents at once are actually useful?
As many as the project has independent pieces of work at that moment, which is two or three most days and five on a large refactor. Past that, agents spend their time waiting for files rather than writing.
Do I need a branch per agent?
No, and that is the point. A branch per agent moves the conflict to the merge, when nobody remembers the reasoning any more. A shared directory with a claim settles the conflict the second it happens.
What happens when two agents want the same file?
The second one waits. It shows on screen, so you can hand it something else immediately instead of finding out an hour later.