兩個編碼智能體應否同時編輯同一個檔案？

Say you split a feature between two coding agents: one builds the /orders endpoint, the other builds the page that calls it. Each works on its own branch in its own copy of the repo, so neither can overwrite the other's files while they work.

The collisions come anyway, usually in the same few places: package.json, a lockfile, a migration, a route table.

In the tool I'm building I went for warnings rather than locks. When one agent changes a file the other has already changed, both chats get one line naming the other task. For those shared files the line is firmer and asks the two to agree on one owner first. The agents can also message each other, so they can settle it themselves. And if two branches still conflict when main comes in, the conflict goes back to the agent that made it, with both sides' intent.

The case for a hard lock is that it's simple: the first agent to touch a file owns it until it lands. The cost is that the second agent has to wait or work around a file it needs, and a lock held by an agent that's stuck blocks everyone.

So I'm curious how others handle it. If you run more than one agent on a codebase, do you split the work so files never overlap, let them collide and sort it out at merge time, or lock files? And what do you do about the shared ones, like lockfiles and migrations?

(I'm building Styx, an open-source app that does the warnings part.)