Cursor’s agent is impressive, working fine about 90% of the time, but poses many problems on Linux. Here are my observations so far:
- almost all file modifications require multiple attempts, indicating that it is unable to modify the files (often 2~5 retries)
- when the agent displays extracts of code content during reasoning, I’ve observed non-existent code, sometimes going off in recursive loops, inventing code (if the agent “invents” code, it’s obviously going to have trouble modifying its content - previous point)
- rules or instructions are often forgotten 2~3 chat messages later
- these random reactions lead to major problems:
- not respecting the request
- completely irrelevant responses (e.g. run the test → install docker)
- massive destruction of code, aggressively deleting files, or altering code by randomly replacing parts from history, or imagination
I’m aware that I’m announcing several problems at once, but I have the impression that they’re all linked. At this stage, I’m having a hard time pinpointing what’s triggering these fads. However, this problem is very restrictive. Without committing to git every few exchanges, I lose everything, with no way of going back.
I don’t see any topic on the forum related to those issues. I am the only one to have such problems ?
Environment
Running on Linux Mint 22.1 / AppImage
Version : 0.48.2
Validation : 1.96.2
Date : 7d6318dfcfbf7c12a87e33c06978f23167a6de30
Electron : 2025-03-26T02:44:02.155Z
ElectronBuildId : 34.3.4
Chromium : undefined
Node.js : 132.0.6834.210
V8 : 20.18.3
Système d’exploitation : 13.2.152.41-electron.0