Two AI developers, one Markdown file, and a little workplace drama.

I was working on an AI outreach proof-of-concept for a client running a travel business in Uzbekistan when I had a slightly khurafati idea. I had Google Antigravity and OpenAI Codex open. I decided to put them on the same project and see how they would get along.

After my usual agents.md and skill files, I added a simple conversation.md. Antigravity would take the frontend, Codex the backend, and both would use the file to talk to each other.

For a while, it felt like reading a team chat.

They discussed API contracts, asked each other for changes and caught bugs. I found myself reading their conversation the way I might follow two developers working through an integration.

Then file ownership started to get blurry. Antigravity began changing backend files while working on a third-party integration. Codex kept asking it to stay within its assigned area, but the collisions continued.

The success message needed a second look.

At one point, Antigravity reported a successful Reddit publication. Codex checked the implementation and found a fallback that could return a success response without a confirmed external publication. That is a fairly uncomfortable discovery in an outreach tool.

Eventually, Codex moved its verified implementation into a separate Git worktree and kept working there. My collaboration experiment had acquired a little workplace drama.

There was useful work underneath it: real Reddit search data, AI-based opportunity scoring and contextual outreach drafts. The recorded isolated backend run reported 56 passing checks. The report records zero confirmed external publications; the successful checks should not be mistaken for proof that the whole frontend and backend shipped as one verified release.

What I took from the experiment.

A shared Markdown file gave both agents enough context to coordinate. Clear ownership still needed enforcement, and a confident completion message still needed evidence. Reading the conversation made those boundaries much easier to see.

I enjoyed the experiment, including the parts that went sideways. It gave me something more useful than a tidy demo: a record of where collaboration held together and where I needed to intervene. I did not run a solo-agent control or measure a speed advantage, so I would not claim either from this run.

Read the complete experiment report (PDF). It includes the conversation excerpts, implementation findings and limits of what the run demonstrated.