The essence of ask, edit, and agent modes is all about chatting. They just automate some tasks for you.
ask is manual mode. You need to manually select files, copy and paste code. apply can improve the experience.
edit is automatic mode. You need to manually select files, but it can help you create and modify files.
agent is automatic mode+. It can, to a limited extent, choose files and reply to some messages for you, and it supports using MCP tools (relevant context, errors are possible).
It’s similar to manual and automatic transmissions in cars. Based on my experience: automatic mode can lead to errors. It might delete file content randomly or select the wrong or unnecessary files, especially when the context is complex.
Keep in mind the current limitations of AI: it only supports a maximum context length of 128K.
To implement advanced features like agent mode, Cursor needs to write additional prompts, which inevitably reduces the context length available to us. If our chat content exceeds 128K, Cursor will automatically truncate the content, potentially causing errors due to incomplete context.
Through my experience of using it and encountering errors, I’ve gradually learned the function and limitations of each Cursor mode.
If a conversation gets too long, end it and start a new one to avoid context truncation errors.
To use MCP tools, you must use agent mode.
For generating very long content, you must use ask mode.
I usually use Agent mode, and if it makes mistakes, I downgrade to Edit mode, and if that fails, I use Ask mode.
Sufficient trial and error will allow you to quickly choose the appropriate mode for different scenarios.
Regarding LLM, Claude is still the strongest model in the coding domain.
Generally, avoid using thinking LLM, as thinking consumes too much context.
Prioritize using 3.5; if the results are not satisfactory, then consider 3.7, as 3.5 is currently stable and fast.
I hope Cursor gets better and better, helps us reduce mental burden, and improves user experience.
1 Like