Agent chats shows dots like it is working when I stopped it… but not only that I stoped it, I had to do that couple of times… and it won’t stop executing. Like some background tasks are still there that wake-up the parent agent that then just keeps going!
Also seems this new version is pushing for verification in browser almost always, which is annoying and not needed every time. Then stopping nightmare happens when you actually want to stop it running these validations.
Seems new Cursor versions wanted to inherit from Grok bot these “parallel” executions with chat… but Cursor should not be Grok bot.
Also models keep switching to Anthropic sometimes without me noticing and doing it on my own.
Maybe the problem is going to a chat that has a model that I used… and then creating a new chat from that chat… anyway. this model switching on auto is dangerous as it drained my tokens without me choosing it (as I mostly do work in Grok 4.6 now)
Steps to Reproduce
Try to stop any browser validation that is happening
Hi @firewalker Thanks for the post. Three separate things are tangled together here, I’ll tackle them one-by-one:
1. Stop taking several tries
You shouldn’t need to press Stop repeatedly. Stop ends the current turn, but background work the agent already started keeps running, and when it finishes it wakes the agent back up. That matches your “background tasks are still there that wake-up the parent agent” exactly, and it’s an issue we’re tracking. That’s also likely your loop: a browser check runs in the background, you stop, it completes, the agent wakes and checks again. If a Working pill appears above the input, opening it and using Stop All is the closest thing to a real stop today. It isn’t fully reliable for local subagents yet, so tell us if that does nothing either.
2. Browser checks firing almost every time
Recent versions intentionally ask Grok 4.6 to verify web changes in the browser. The frequency clearly isn’t right for every workflow, and I’ve passed that along. For a fast first pass, add this to a prompt or a rule: Do not verify changes in the browser or take screenshots unless I explicitly ask. That competes with the built-in instruction rather than disabling it, so for a hard stop, turn off Browser Automation under Cursor Settings > Browser & Network.
3. The Anthropic models
Most of the Anthropic usage here is expected. Some browser automation tasks are pinned to specific models rather than following the model you pick in chat, so Anthropic can appear in your usage even while you’re working in Grok 4.6.