DEV Community

Richard Smith
Richard Smith

Posted on

I Started Listening to What My AI Agents Hate About My Code

A few months ago I noticed something strange. My AI agents kept getting stuck on the same edge cases, producing inconsistent outputs, and sometimes just refusing to work with certain parts of my codebase. I initially thought this was a limitation of the model, but then it clicked—what if these "failures" were actually feedback?

I started tracking the complaints. Not the errors, but the patterns. Which parts of my product made agents work harder than necessary? Where did they get confused? What assumptions did they keep breaking?

The answers were genuinely useful. I found entire features that seemed fine to me but created friction for any automated process. I discovered inconsistent naming conventions that confused both humans and models. I even found some legacy decisions that made perfect sense five years ago but now actively hindered any kind of automation.

Now I'm thinking about this as a product angle. If you build tools that systematically collect and surface this kind of agent feedback, developers could use it to improve their codebases—or even their products—with real, grounded data instead of guesswork.

Anyone else tracking agent friction as a signal for improvement?

Top comments (0)