Improve verification

What feature would you like to see?

First off, I find Codex to be an excellent tool in combination with the max model. It’s the first tool, that seems like it actually understands architecture and implementation properly.

With the apply_patch now integrated into the model, and the recent improvements, I find Codex can solve almost all the assignments I give it. From small POC projects to huge enterprise systems. But not always on the first try (and I think that’s expected).

For ML/exploratory work, I have added tests and just told Codex to improve model architecture until a certain MAPE was achieved in the test— and it worked for 14 hours :-)

BUT, when I develop classic frontend/backend applications, Codex doesn’t have this ability to verify its work the same way.

A typical workflow it:

1) ask codex to implement feature. Codex edits and build frontend/backend
2) verify in the frontend
3) adjust in Codex
4) verify in the frontend
5) adjust in Codex
6) verify => finally ok. Git commit

I think a missing piece to Codex/model interaction is to improve how Codex verify output in browsers/applications after edits. That would close the loop. I have tried with Playwright and Chrome mcp but it’s not effective. Codex doesn’t seem trained to understand and use these tools and it often fails and say everything is ok, when it’s not (like console errors show or an element is missing)

Imagine if Codex could much more reliable do computer use and check the DOM or using images. It would greatly improve its agentic abilities, but I think it has to be deeply integrated into the model and Codex CLI.

Additional information

_No response_