User post

Dot tests what happens when two agents edit one file

Yeemio says he asked his Dot, his AI assistant, to run a small test on its own cloud computer: what happens when two agents work on the same file. He reports that a lease let the agents take turns and keep both edits, but that bypassing the lease and writing an old view overwrote the file, after which the delivery packet failed its file-hash check.

What it is

Yeemio (@yeemio) posted that he asked his Dot, his AI assistant, to run a small test on its own cloud computer: what happens when two agents work on the same file. He says the machine runs Debian, and that two real agent sessions shared one directory and used a copy of an existing OwlRunKit 0.24.1 npm installation. According to the post, Agent A claimed the file with a lease. While it was active, B's request for the same file was rejected. After A released it, B reread the latest contents, acquired the lease and made its change. Both edits remained. He says that on different files, both leases were active and the agents' work windows overlapped, and that they deliberately held each open for two seconds to observe the overlap, not to measure speed. He also says they deliberately bypassed the lease and wrote an old view directly. The overwrite succeeded. The original delivery packet then failed its file-hash check. He writes that a lease coordinates agents that use it, that it isn't an OS file lock, and that the packet detects the overwrite after it happens. This page restates that user post in English. It is not an OpenAI statement and not an independently verified case study.

Evidence

User post on X by Yeemio (@yeemio). Not an OpenAI statement and not an independent case study.

Caveats

  • The poster’s X post only. This page does not add details the post does not state.
  • Not a verified independent case study. Not an OpenAI statement.
  • Any outcomes or tool comparisons are the poster’s description only.

This page restates a user’s X post. It is not an OpenAI statement and not an independent case study.