Simple tasks seems to be taking an extra long time, while stuck on the first step after a prompt when agents are set to Auto. I switched to Opus and Sonnet 4.6 earlier in the day, but that eats through API usage. Now, Composer 2 also has the same behavior.
Steps to Reproduce
Stop prompt, change model, resubmit prompt
Expected Behavior
AI should assess and implement fixes quickly and efficiently.
Hey, thanks for the report with the Request ID and version. This looks familiar, we’ve gotten a cluster of similar reports over the last few days about Auto/Composer 2 being slow:
From your request, we can see a short stall in the stream, but the request on the server finished without errors. This looks like latency, not throttling.
We’re already tracking a performance regression in agent chats internally, but we don’t have a clear ETA for a fix yet. If it gets better or worse after the next updates, let me know and I’ll update the thread.
Workarounds for now:
Cancel + resubmit when it hangs on the first step, like you did.
For long sessions, start a new chat once the context gets really large.
Hey @BobbyThirtyTwo coming back to this thread. Since your report when you were on 3.2.16, we’ve shipped a few releases with agent performance improvements, so things should look different now.
If Auto or Composer 2 is still getting stuck on the first step after your prompt, please update Cursor to the latest version. If it happens again, start a new thread with the latest details and a fresh Request ID. Privacy Mode needs to be off. We’ll take a look with clean, up-to-date info.
For now, I’m going to close this thread. If anything comes up, post in a new one and we’ll help you out.