Idea

Take the browser control story seriously for a week. Instead of cherry-picking easy pages, give the agent popups, delayed content, sticky headers, and forms that shift underneath it.

Status

Some tasks worked. The interesting part was how often near-success still meant failure because the last ten percent was timing, not logic.

Problems Faced

A lot of browser automation is really state management disguised as clicking.
The agent needed stronger guardrails for retries and verification.

Techniques That Helped

Assert the page state after each meaningful action.
Prefer slower, explicit checks over optimistic progress.

Lessons Learned

Agents do not just need tools. They need a disciplined definition of done.