Cursor auto agent is demonstrating prior version logic and bugs similar to that of GPT 5.2 and earlier. I would like to know what verstion logic is available for the cursor auto agent because the output is significantly changed and the drift is beyond measure. Agents will not follow basic instructions or markdowns as prescribed and drift if more than two sentences.
Steps to Reproduce
note that hyphens show up in most text generation, context limiting and inability to handle complicated or concurrent logic.
Expected Behavior
I expect this to work like it did prior, i fear that cursor rolled the agent logic back to save a penny and water down the value of the auto agents.
Hey, thanks for the report. A few users have mentioned something similar with Auto mode, and the team is aware.
Auto mode dynamically picks a model for the task, and that choice isn’t always ideal, especially for complex multi-step tasks. A couple things that might help:
Try switching to Premium mode next to Auto in the model selector. It prioritizes quality.
Or pick a specific model manually, like Claude Sonnet, Claude Opus, or GPT-5 for harder tasks.
If you want us to look into a specific case, send the Request ID (three dots in the top right of the chat > Copy Request ID). With that, we can check which model was picked and what went wrong.
I got the same issue, and also mentioned several places.
Agents are down all over.
Even premium agents, either take a really long time, or most of the times time out.
Request ID: 6d5946d3-eae3-4282-bc92-f279d8aea05a
{"error":"ERROR_EXTENSION_HOST_TIMEOUT","details":{"title":"Agent Stream Start Timeout","detail":"The extension host did not respond in time. Please reload the window to continue.","isRetryable":false,"shouldShowImmediateError":true,"additionalInfo":{},"buttons":[{"label":"Reload Window","reloadWindow":{}}],"planChoices":[]}}
Agent Stream Start Timeout [deadline_exceeded]
ConnectError: [deadline_exceeded] Agent Stream Start Timeout
at B (vscode-file://vscode-app/d:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:37752:74001)
at vscode-file://vscode-app/d:/Program%20Files/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:37752:74271
@Casey_Salmon If the issue still happens on Premium, try manually selecting a specific model Claude Sonnet 4.6 or GPT-5. This will fully disable automatic routing. If the quality is still bad with a specific model, please share the Request ID for one of the affected requests three dots in the top right of the chat > Copy Request ID. With that, we can check what’s happening on our side.
@master3395 Your Agent Stream Start Timeout / extension host issue is a separate bug and not related to Auto mode quality. There are similar reports here: BUG: Waiting for extension host. Try:
Disable extensions run cursor --disable-extensions and check if the issue still reproduces
If you’re on Nightly, try switching to Stable since Nightly (3.1.0-pre) can be less stable
If that doesn’t help, it’s best to create a separate thread with your Request ID so we can investigate your case directly.
nah, the answer isnt always switch to a premium model unless your giving free model usage.. the issue is the auto agent. This last patch does seem to have improved things though. I swear, downloading a cursor patch is like a box of chocolates…
Fair point, Premium shouldn’t be the default answer. Glad the latest patch worked better. The routing in Auto mode gets tuned regularly, so noticeable quality swings between versions are unfortunately not rare.
If you see a конкретно bad answer again, send the Request ID (three dots in the top right of the chat > Copy Request ID). I’ll check which model it went to and what happened. This really helps us catch cases where Auto picks the wrong model for the task.
