Race Agents on One Task

The Idea

Agent output varies a lot between attempts and between agents. Instead of prompting, reviewing, and re-prompting one agent serially, run the attempts in parallel on isolated branches and pick the winner.

Steps

1. Press ⌘N three times. Same prompt each time, different agent each time (or the same agent three times: variance alone makes racing worth it).
2. Keep Auto-run command on so each agent starts immediately.
3. Write the prompt once and reuse it verbatim. End it with a self-report so comparison is fast:

Add rate limiting to the public API endpoints. Use the middleware pattern
already in this codebase. In your final message, list the files you changed
and any tradeoffs you made.

4. Do something else. Notification sounds tell you when each agent finishes.
5. Open each workspace's Changes tab. Focus mode (target icon) makes stepping through a diff file-by-file fast.
6. Ship the winner: stage, commit, push, Create PR from the diff viewer.
7. Before deleting the losers, skim their diffs for ideas the winner missed. Select and copy any lines worth keeping.
8. Delete the losing workspaces.

Picking a Winner Quickly

• Read the final messages first; disqualify anything that misunderstood the task before reading diffs
• Prefer the smallest diff that fully does the job
• Run the test suite in the winner's terminal before you push

Variations

• Two agents, two approaches: same agent, but prompt one toward each design ("use middleware" vs "use a decorator") when you're unsure which is right
• Best-of-N on hard bugs: for a bug that has resisted one agent, three fresh attempts beat continuing a stuck session
• From the CLI: write the prompt to prompt.md, then start the whole race in one paste — one workspace per agent, same prompt each:

for agent in claude codex opencode; do
  superset workspaces create --project <projectId> \
    --name "race-$agent" --branch "race/$agent" --local \
    --agent "$agent" --prompt "$(cat prompt.md)"
done