I notice the last day something major has happened. I exclusively use “auto” and it has been very consistent for months, or maybe even a year. Starting sometime this week something has definitely changed for the worse.
I have a few concurrent and very different projects I work on daily for different clients and all of them are feeling nerfed. The agent is responding very differently?
All of a sudden the agent starts randomly restoring back to abandon approaches in the code. Even after repeatedly telling it to stop going down a certain path it continues to re-add the files and approach.
It started running applications to launch another project. It doesn’t even prompt to do this but as if it is compelled to deploy the project. Very strange as it never did this over the weeks of working on it.
It really has a smell of something very different than the consistent and excellent results I had for months.
This reminds me of the early days of Cursor where every couple of weeks something fundamental changed and the results went from one-shots to reverting entire commits.
I hope this is temporary and to be honest I sort of expected a Grok model to be inserted into “auto” - please no
Hi @eatatjoes Thank you for the forum post and I apologize for the unexpected behavior of auto mode here. I checked your account. Your Cursor version and settings haven’t changed, but two things stand out:
Auto’s underlying model selection isn’t fixed. It can shifts over time as models are added and evaluated, so the mix you’re getting this week is slightly different from what you had a few weeks ago. Cursor doesn’t share the exact models that Auto is running on.
On Sunday evening (your time), a run of your Auto requests hit a capacity limit and errored repeatedly for about 25 minutes before recovering. When that happens partway through a task, the agent can lose its place and redo or revert work, which lines up with what you described about it going back to approaches you’d already abandoned.
If consistency matters more to you than letting Auto choose, the quickest fix is to pin a specific model in the model picker for the projects you work on daily.
If you see this again, a Request ID from the affected turn would help. Open the conversation, click the “…” menu, and choose Copy Request ID. One note: Privacy Mode is on for your account, so we can see that a request failed, but not what the agent was doing. If you’re willing, temporarily enabling Share Data before reproducing gives us a much clearer picture. Details here: Reporting a bug | Cursor Docs
