Rules for running several coding agents on the same repos
A mixed setup ran Claude Code as the coordinator, Codex for most work, and OpenClaw plus an ACP-connected agent for specialized tasks across the same repos for months. The main lesson is that different agents fail in different ways, so the workflow needs clear separation and clear stop rules. Any agent that writes files should get its own .
When two agents edit the same checkout, they create conflicts that neither can predict. Before reset, checkout, stash, or branch changes in a shared repo, every unfinished change should be saved as a labeled , whether it came from a person or an agent. Each task brief should include a failure path: after three stuck attempts, stop, revert, and report.
Review should happen once, and the reviewer should not be the builder; its job is to challenge the work rather than approve it by default. Approval can also include a short quiz on the diff so the human actually checks the change before accepting it.
Key points
- Give every file-writing agent its own .
- Avoid letting two agents edit the same checkout at the same time.
- Save unfinished work as a labeled before reset, checkout, stash, or branch changes.
- Tell agents to stop, revert, and report after three failed attempts.
- Use a reviewer that did not build the change, and make it look for problems.