When should a browser agent stop and ask for help?

While building rea-captcha, I found the hard part wasn’t detecting a CAPTCHA. It was deciding what counts as enough evidence to let the original Browser Use task continue.

My current rule is simple:

• keep the browser session the agent already owns

• use bounded automation only on owned pages or provider test keys

• pause unknown and production challenges for a person

• after handoff, re-read the current page; a click or token alone is not success

For people running Browser Use or Playwright agents: do you keep the session open, restart the task, or let the agent retry? I’d especially value failure cases where the task keeps looping after the CAPTCHA is gone.

Implementation and four controlled demos: https://github.com/masakaai/rea-...