Getting worse results from browser tasks provided via SDK rather than via cloud UI

[j-petty]

I'm not sure how this is happening but it's very consistent. When creating a task via the SDK I have noticed the agent usually often gets stuck on tasks like submitting a form. However if I provide the same prompt via the https://cloud.browser-use.com/ UI the agent performs much better.

I've confirmed I'm using the same model for both, and have tried several models. This happens regardless of the model used.

I've noticed the live preview is rendered differently for tasks submitted via the UI vs SDK, suggesting there is a setting that is perhaps not being surfaced, for example maybe they're using different default browser profiles. But given I'm not specifying a profile in either case this is not clear.

I'd be interested in knowing what the difference would be, I'd expect them to call the same API in the background. Keen to hear if anyone else has found similar.

[LarsenCundric]

Hey there! So the SDK uses the new V2 API which uses our in-house stealth browser (its in beta). This is the only difference and might affect performance... We are actively working improving it tho.

Could you maybe provide me some example runs so I can look into them? You can just email me on larsen@browser-use.com.