I think we're underestimating how much control coding agents actually need

AI coding agents are getting really good at writing code. You can describe a feature, give the agent access to your repository, and let it modify files, run commands, install packages, write tests, and keep iterating until something works. That's impressive. But I think there's another problem that doesn't get enough attention: What stops the agent from solving the problem in a way we don't actually want? For example, an agent might: generate code that works but violates your architecture skip tests because the task appears simple introduce unnecessary dependencies change unrelated files optimize for passing tests instead of maintainability make security-sensitive changes without clearly explaining them At that point, I don't think giving the model a better prompt is enough. I'd rather define explicit rules for the agent: What it can change. What conventions it must follow. What tests it must write. What it has to verify before finishing. And, importantly, what exactly it changed and why. The interesting shift for me isn't just: It's: Maybe the next generation of coding agents needs less freedom, not more. How are you handling this in your own coding agents?