After a recent conversation here, I spent a few weeks using agents.
These agents are just as disappointing as what we had before. Except now I waste more time getting bad results, though I’m really impressed by how these agents manage to fuck things up.
My new way of using them is to just go back to writing all the code myself. It’s less of a headache.
Cursor, Copilot agent mode, and Windsurf. The agent modes can search repos, modify code, and run code on their own. I thought Cursor's agent was the best. I did like Windsurf's agent plans, but the actual results weren't good. Copilot's agent has been slow and not good. But basically all of them were like a pretty bad junior engineer - sometimes it would hit the right result, but usually not. The code often looked good but rarely even ran, let alone met requirements. They would frequently break things, I'd fix them, they'd break them again. Most of the time this cycle was slower and more frustrating than just writing the code myself. I tried one or two one-shots on lovable - the design was impressive, but functionality and attention to specs were poor.
I've had the most success with extremely small questions rather than asking the agents to write a lot of code. In those cases, the code is still usually wrong, but it's close or small enough that I can quickly fix it.
Don't get me wrong: I find all of these tools to be really impressive and good enough to be useful. But the improvements and huge productivity gains friends claim they or their workers are getting just aren't materializing for me.
These agents are just as disappointing as what we had before. Except now I waste more time getting bad results, though I’m really impressed by how these agents manage to fuck things up.
My new way of using them is to just go back to writing all the code myself. It’s less of a headache.