Coding agents are getting pretty good at writing code, but they still have no way to verify if they break the UI.

As humans, we rely on visual diffs for that. We open them, scan quickly, and catch obvious regressions. Agents are completely out of that loop.

I’m a co-founder of Argos (visual testing), and I recently shipped a CLI to expose visual diffs in a way an agent can actually use, instead of going through a UI.

Once I wired it into an agent workflow, a few interesting things started happening. The agent started catching obvious regressions. Sometimes it would refuse to approve its own PR. With a good prompt, it even fixed the issue after seeing the diff and iterating.

It’s still rough and not reliable enough to trust on its own. A lot depends on how well the agent understands the codebase. In local tests, it sometimes gets stuck in loops and burns through tokens .

Giving agents “eyes” on UI changes might be an interesting feedback loop for more autonomous dev agents in the future.