Complaint Regarding Cursor AI Agent Performance

Discussion
1

I am writing to express my significant dissatisfaction with the performance of the Cursor AI agent. Frankly, I find the agent to be quite ineffective and frustrating to work with for several reasons

  1. Repetitive and Misleading Modification Process: The most glaring issue is the agent’s tendency to repeatedly state “Finally…” or similar concluding phrases when suggesting code modifications, only to propose further changes immediately after. This happens numerous times, making the editing process unnecessarily long and confusing. It gives a false sense of completion.
  2. Drastic Performance Drop with Non-Anthropic Models: The agent’s ability to utilize tools and generate accurate code is severely lacking, especially when not using Anthropic’s models. While Anthropic models might perform adequately, other models exhibit a catastrophic drop in performance, frequently failing at tool calls and code generation. This limitation significantly hinders its usefulness and flexibility.
  3. Lack of Transparency and Resource Consumption Without Results: There is a concerning lack of transparency regarding token usage and the agent’s internal decision-making process. Furthermore, the agent often consumes my paid quota (“$0.04” or fast request allocations) without actually delivering meaningful or correct code modifications. It feels like I am paying for the agent to attempt modifications rather than successfully implementing them. It frequently fails to apply the requested changes correctly, wasting both time and resources.
  4. Inconsistent Performance Compared to Other Environments: My experience contrasts sharply with using models in other environments, such as “roo code”. In “roo code”, various models demonstrated smooth and optimal performance for coding tasks. This suggests the performance issues lie specifically within the Cursor integration, not necessarily the underlying models themselves.

Overall, the current state of the Cursor AI agent is disappointing. It struggles with basic modification loops, exhibits drastic performance differences between models, lacks transparency, fails reliably with certain configurations, and consumes resources without providing commensurate value. I urge the Cursor team to address these critical performance and transparency issues.

4 Likes