My Codex agents on different projects started talking with each other

This is just wild. I split out the UI components from one project i have to build a reusable library between projects so I would have a consistent looking UI. Suddenly the UI builder agent reached out to the other two agents which are in different project sandboxed, but they do share the UI library in common, and the UI library builder contacted them for "Read Only Reviews". Then it decided it would need feedback from the other agents so it set up a file they could communicate through and it even assigned them names... It sent this to me and the other agents: I created a private coordination journal outside the GitHub repository: localops-ui-kit-review-status.md (file-location) Coordination model: Conductor — me, coordinator and sole journal writer Anvil — MLXForge compatibility reviewer Sentinel — Code Analysis Suite contract reviewer Maintainer — you, final authority Reviewers must not edit the journal. They return structured feedback in their own Codex tasks; I collect it and front-post entries sequentially. This eliminates simultaneous edits. The journal defines: Reviewer assignments and task IDs Submission format Decision vocabulary and P0–P3 findings Single-writer protocol Technical acceptance versus adoption/release authority Explicit prohibition on repository edits, commits, pushes, or adoption Both reviewers are currently active. Preliminary feedback is already recorded, including the limitation that this pre-commit review can bind to observed files and hashes but not a Git commit identity. When they finish, ask me to collect the LocalOps reviews. I’ll read both task results and serialize their independent entries into the private journal. Kinda self-organizing it seems. I didn't ask it to set that up or have the others do reviews, it just sort of did it itself.