Where does the bug appear (feature/product)?
Cursor IDE
Describe the Bug
When using Agent mode in Cursor IDE, selecting the “Auto” model triggers a fatal error (exceeded max retries) and renders the Agent function unusable. The Agent mode only works normally when manually switching to the more costly Opus 4.6 or Sonnet 4.5 models, which increases usage costs and defeats the purpose of the Auto model.
Steps to Reproduce
-
Open Cursor IDE (latest version) on Windows 11.
-
Navigate to the model selection dropdown menu and switch Cursor to Agent mode.
-
Select the “Auto” model from the model list (instead of Opus 4.6/Sonnet 4.5).
-
Attempt to perform any Agent-related operation (e.g., code generation, code refactoring, or natural language QA via Agent).
-
Observe the immediate error popup/console log with the “exceeded max retries” message.
Expected Behavior
The Auto model should be invoked normally in Agent mode: it should automatically select the most appropriate (and cost-efficient) model (e.g., Haiku/Sonnet/Opus) based on the complexity of the Agent task, run without retries/exceptions, and maintain full functionality of the Agent feature.
Screenshots / Screen Recordings
No screenshots attached (error is text-based in logs), but the full error stack is provided below.
Operating System
Windows 11
Version Information
IDE:
Version: Latest stable version (exact version copied from About Cursor: 2.6.12 (user setup))
VSCode Version: 1.105.1
Commit: [1917e900a0c4b0111dc7975777cfff60853059d0]
CLI:
Not applicable (issue occurs in IDE, not CLI)
For AI issues: which model did you use?
Model name: Auto (error occurs); Opus 4.6 / Sonnet 4.5 (work normally)
For AI issues: add Request ID with privacy disabled
Request ID: 3c03d011-4f43-4458-9284-f15402881eac
Additional Information
Full error stack trace:
plaintext
[internal] Failed to run step, exceeded max retries
_he: [internal] Failed to run step, exceeded max retries
at jpA (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:32115:44883)
at WpA (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:32115:43786)
at ZpA (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:32116:5088)
at Ool.run (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:32116:9098)
at async s$o.runAgentLoop (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:44360:8423)
at async zOl.streamFromAgentBackend (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:44408:8884)
at async zOl.getAgentStreamResponse (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:44408:9837)
at async yLe.submitChatMaybeAbortCurrent (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:32182:15752)
at async Gs (vscode-file://vscode-app/f:/cursor/resources/app/out/vs/workbench/workbench.desktop.main.js:43415:4781)
-
The issue is reproducible 100% of the time when following the steps above.
-
Using Opus 4.6/Sonnet 4.5 bypasses the error but incurs higher AI usage costs.
Does this stop you from using Cursor?
Sometimes - I can sometimes use Cursor (by switching to more expensive models), but the Auto model is completely unusable in Agent mode.