Agentic models becoming progressively worse

Discussion
1

Ive been using Cursor with Agentic mode for about 5 months now. I have to say since about late November into December its just slipping more and more. I don’t know if anyone else experiences this? Just today, I tried to break this prompt down into small pieces for it:

I need a comprehensive, end-to-end overview of this entire project. That means:

1. Inspect every component in the codebase and explain:
   - What it does
   - Which files are involved
   - How it interacts with other parts

2. Examine the backend database in detail:
   - Describe its schema, indexes, and design
   - Use the existing script in the codebase to confirm or gather this information

3. Stop after 10 files are analysed and await instructions. If you’re unsure about anything, ask for clarification before making assumptions.

4. Do NOT skip or gloss over any major sections. Your report must provide a holistic understanding of how everything fits together.

Make sure you cover the entire codebase and database architecture, not just isolated pieces.

It did 3 then stopped, then guessed what the others did. I told it to read its instructions again and it did 4 then guessed the rest. These are by no means complicated tasks and its abysmal at them. Not sure whats going on but its becoming worse over time. Used to be able to give it as task and it would execute it more or less flawlessly.

Does anyone have any tips on how to instruct the LLM on how to do a full analysis?

Its getting to the point where its not fun to use anymore. Im spending more time trying to get the agent to understand the task and debugging its attempt at coding than actually getting stuff done. It used to have understanding and be able to identify relationships in the code but now its losing chunks of IQ points month on month.

2

I totally get the frustration with the analysis performance. Let me share some tips that should help get better results:

  1. Break down the analysis into smaller chunks using @Folders instead of @Codebase. This helps prevent the model from getting overwhelmed

  2. Enable Codebase Indexing if you haven’t - this dramatically improves code relationship understanding

  3. Create a .cursor/rules file with specific instructions about how you want analysis performed. The model will follow these consistently

  4. Use Agent’s continue feature when it stops - each continuation lets it make 25 more tool calls

For the database analysis specifically, you might want to run that as a separate task after the codebase overview since it requires different tooling

These tweaks should help get closer to the performance you were seeing before.