My introduction to agentic engineering was at Slack in December 2023 with Cursor. It blew my mind — I felt like I could talk to my code, which in retrospect was this weird, uncanny valley-esque experience. Now I've grown used to it. Cursor, like me, changed along the way. Cursor started as this ambitious IDE that was far more equipped for the multitude of models that were appearing on the market. A reminder for those that don't know or have forgotten, Microsoft was strategically speaking making it very difficult to bring in models other than theirs into the VSCode agent pane.
That's all changed. Cursor now has two modes: agent and IDE. The IDE still has an agent interface but it's not the main show. The agent interface mode dominates the window and is more like a somewhat more advanced, but simultaneously less advanced Claude Code. The real value to Cursor these days is actually a model! My, my how things have changed in such a short time, right? Composer 2.5 is a joy to work with and on the extreme end of being economical. It basically costs nothing and has similar performance to the Opus models.
With a Claude Max plan I found myself using Cursor less and less. For context, I justify most of my AI spend alongside my consulting business. I use Claude to manage the business, write code, stay organized – all the things. Cursor was always just a coding tool, but it was difficult to use alongside Claude Code. There never was that deep of an integration between the two, although I do believe that one exists. Eventually I questioned my $60/month spend and I started paying attention every time Zed made HackerNews for another press release. I downloaded it, configured it to work alongside Claude, and I was surprised to find out that it was basically exactly what I was looking for.
What Claude doesn't give me is an in the weeds view of my code. It is not an editor and does not even try to be. The best it can give me is diffs and they're not even that great – which is fine. I want Anthropic to focus on providing me stable service for my frontier model subscription not other side projects. Zed has a window that talks directly to Claude Desktop and Claude Code. You might be asking yourself at this point, "Why are you even reading the code?" I don't always get in the weeds with Claude when writing code to be fairly honest. Often enough I'm just enhancing a feature, fixing a bug, or whathaveyou. I tend to shore up my various layers of testing such that we're working in what amounts to layered TDD so I have strong confidence that what I tell Claude I want it to do is getting done.
I do, however, want to see the code when crafting entirely new systems. Claude can sometimes get inventive or cut corners, violating established repository patterns in favor of whatever-the-fuck Claude wants to do in the moment. This is where working with Zed I think really pays off; I can have my in-the-weeds view with Zed while enjoying the use of my Claude Max subscription. I've saved $60/month and meaningfully improved my workflow.
Cursor isn't a bad product; I actually like it quite a bit and use it every day at work. I just have better options at home.