Updated 8 October 2026

How to race several coding agents on one prompt

Write the prompt once, give every attempt its own git worktree and branch cut from the same commit, start one agent in each with that same prompt, then compare the diffs and merge the one you like. Delete the rest. Because the starting point and the prompt are identical, the only thing that differs is the agent, or the model it runs on.

This is how you compare agents on your own code instead of on someone else's benchmark. You can do it by hand with git, and this page shows how. Further down it shows what Agency's race does for you. Disclosure: Agency is ours, and this site is its home.

Why race instead of picking one agent

Two agents given the same task rarely produce the same change. One edits three files and one edits eight; one adds a test and one does not. You find that out only by reading both. A race makes the comparison cheap enough to do on a real task: you pay for a few attempts in parallel and read the diffs side by side, instead of running one attempt, disliking it, and starting over.

Race agents by hand

The commands assume a repository called myapp with a main branch, and three attempts: Claude Code, Codex and Gemini CLI.

1. Cut one worktree per attempt from the same commit

   Run these from inside myapp. Naming the same start point for all three is what makes the race fair.

   git worktree add -b race/claude ../myapp-claude main
   git worktree add -b race/codex ../myapp-codex main
   git worktree add -b race/gemini ../myapp-gemini main

2. Set each one up the same way

   A worktree is a fresh checkout, so install dependencies and copy untracked files such as .env into each. If one attempt starts with a working environment and another does not, you are comparing setup, not agents.

3. Send the same prompt to each agent

   Save the prompt to a file so every attempt gets exactly the same words. In tmux, each agent gets a window, and you paste the prompt into each:

   tmux new-session -d -s race -c "$PWD/../myapp-claude" claude
   tmux new-window -t race -c "$PWD/../myapp-codex" codex
   tmux new-window -t race -c "$PWD/../myapp-gemini" gemini
   tmux attach -t race

   Most agent CLIs also accept the prompt as an argument when they start, which removes the pasting. Check each CLI's own help, because the flag differs.

4. Compare what each one changed

   When an agent says it is done, read its change against the commit it started from. Start with the size, then read the diffs:

   git diff --stat main...race/claude
   git diff --stat main...race/codex
   git diff main...race/claude

   Run the same test command in each worktree too. A diff that reads well and fails the tests has not won.

5. Merge the winner and delete the others

   git merge race/claude
   git worktree remove --force ../myapp-codex
   git worktree remove --force ../myapp-gemini
   git worktree remove ../myapp-claude
   git branch -D race/codex race/gemini
   git branch -d race/claude

   --force and -D are there because the losing attempts have uncommitted or unmerged work that git will not discard otherwise. That is the point of the step, but check the diffs first.

What to look at when you pick a winner

• Whether the tests pass in that worktree, not whether the agent says they do
• How much it changed: a smaller diff that does the job is easier to review and to revert
• Whether it touched files the task never needed
• Whether it followed the conventions around the code it edited

You can also take pieces from more than one attempt. Merge the best one, then git cherry-pick a commit or copy a hunk from another.

What gets hard by hand

• Three worktrees, three installs and three windows for one question
• Pasting the same prompt three times without it drifting
• Telling which attempt has finished and which is waiting on you
• Remembering which branch belongs to which agent when you come back to it
• Removing the losing worktrees and branches without removing the winner's

Doing the same with Agency's race

Agency is a free, open-source desktop app for macOS and Linux that runs agent CLIs in real terminals, each in its own worktree. Its race is the steps above behind one dialog. Open the add-agent menu in a project and choose Race agents.

A race needs at least two attempts, and it needs a git repository, because every attempt is a branch. For a project that is a plain folder, the race option is hidden. You can also start a race from an issue, so the issue's text is the prompt.

An honest note on evidence: as of 8 October 2026 this page has no race from our own history to show you. Agency's own repository has merged more than 200 agent/ branches, but a merge commit names only the winning branch, and the losing attempts are usually deleted after it. Nothing left in the history says which merges came out of a race. We would rather say that than invent a scoreboard.

Who should not use it: anyone on Windows or an Intel Mac, which Agency does not support yet, and anyone who only ever needs one agent on a task. Racing multiplies usage, since every attempt spends its own plan or API tokens, so reserve it for work where the choice matters. The by-hand steps above are enough if you race twice a month.

Questions

Does racing cost more?

Yes, roughly in proportion to the number of attempts, because each agent does the full task and bills against its own plan or API key. Race the tasks that are hard to specify or costly to get wrong, not the ones you could review in a minute.

Can I race one agent against itself?

Yes, on different models, as long as the agent's CLI can be told a model at launch. In Agency this is the + model button on an agent that is ticked in the race dialog. By hand, pass each CLI's own model flag in its window.

Can I race on top of work I have not merged?

Start the attempts from whichever branch holds that work. By hand, replace main in the git worktree add commands; in Agency, change the branch the dialog starts from.

New to worktrees? Start with the guide to running coding agents in parallel.