whats your actual system for two agents on one repo, because mine just failed

I finally tried the thing everyone talks about, claude code on the backend task and codex on the frontend task, same repo, same afternoon, and it went fine for about two hours and then it went extremely not fine.

Codex refactored a types file that claude was mid-way through using. Neither of them errored. They just kept going, each convinced the file was what they last saw, and I merged the first branch without noticing the second one had quietly built on a version that no longer existed. Took me most of the next morning to untangle which changes were real.

The dumb part is I knew this could happen and did it anyway because the speed is genuinely addictive when it works.

What ive tried since. Separate worktrees, which fixes the overwriting but now neither agent knows what the other one did, so they duplicate stuff. A little shared markdown file both agents are told to update, which works until one forgets. And ive been running them inside coldtea ai this week, which does the coordination part properly since thats kind of its whole thing, though its mac only and the qa agent it ships with flags more than i want on a prototype repo.

honestly none of these feel like the obvious winner yet.

so what are you all actually doing? one agent at a time and accepting the speed loss? worktrees and eating the duplication? or is there a setup i havent found