Hi,
Thanks for logging this! Since Computer Use is still in preview, model behavior differences like this are expected. Here are a few things worth trying:
1. **Refine your instructions** — Instead of a general "scroll and extract" prompt, break it down step by step: open the URL, read the first visible item, scroll using Page Down, repeat up to N times, and stop after 3 failed scroll attempts. Explicit steps help Claude avoid looping.
2. **Enable Human Supervision** — This setting pauses execution and notifies a reviewer via Outlook when the agent gets stuck, preventing infinite loops.
3. **Switch to a dedicated machine** — If you're using the hosted browser, try configuring a dedicated Windows VM instead. The hosted browser can be throttled and may behave less predictably.
4. **Stick with OpenAI CUA for now** — Since it already works for your scenario, it's the most pragmatic choice while this gets resolved.
For reference: [Microsoft Docs – Computer Use Troubleshooting](https://learn.microsoft.com/en-us/microsoft-copilot-studio/troubleshooting-computer-use)
Hope this helps!
Paolo
✅Did this solve your issue? → Accept as Solution 👍 Partially helpful? → Click "Yes" on "Was this reply helpful?" or drop a Like!
Want more tips on Power Platform & AI? Follow me here:
Thank you for your reply! To ensure a great experience for everyone, your content is awaiting approval by our Community Managers. Please check back later.