[agotterer]

I’m excited about the space and intend to keep an eye on you guys. I actually gave the opened source version of browser-use a try last week and ran into two problems:

The first, it refused to correctly load the browser tab and would get stuck in a loop trying. I was able to manually override this behavior for the purpose of prototyping.

The second, it hallucinated form input values. I provided it strict instructions on how to fill out a form and when it didn’t know what to do with an address field, it just wrote 123 Main St instead of not being able to complete the form.

The thing I really want and haven’t found in any of the browser agents I’ve tried, is a feedback loop. I don’t personally know what the final format looks like. But I want to know that what I expected to happen inside the browser, actually happened, and I want it to be verifiable. Otherwise I feel like I'm sending request into a black hole.

[gregpr07]

Yeah - one version had that weird bug - we fixed it already (I think, super happy to tell me how to replicate if not).

Hmm really? Maybe you could use the sensitive data api to make it more deterministic? https://docs.browser-use.com/customize/sensitive-data

How would you imagine a perfect feedback loop?