Since developing a RSS browser extension with Cursor earlier this year, I’ve always considered Cursor to be the leader in AI programming, far ahead of alternatives. I’ve previously shared my experience and tips for using Cursor as a recommendation to others.
Recently, I wanted to iterate on this extension’s functionality. After Cursor upgraded from version 0.45 to 0.48, I suddenly felt extremely uncomfortable using it. It’s no longer precise in solving and analyzing problems - it feels like a lottery now. Even using the Claude 3.7 model yields equally poor results. The new version of Cursor seems to have lost its adherence to rules. The Agent mode appears to prioritize its own rules rather than following user instructions.
On the surface, Agent mode can automatically fix problems and continuously loop to fix error issues, but in reality, it feels out of control. You don’t know what it’s repeatedly running, fixing, or tinkering with. It feels like it’s constantly guessing at solutions to problems, continuously looking for possible issues and fixing errors. During this process, it might automatically change approaches, giving you the agonizing feeling that your code is being messed up and arbitrarily modified.
Cursor was better before these iterations - it seems to get more confusing with each update. This suddenly creates opportunities for other AI programming tools. Cursor may have been led astray by various Agent implementations. These Agents appear intelligent on the surface but are actually style over substance - sacrificing precision, acting on their own initiative, requiring frequent rework, and greatly reducing efficiency.