Last week I caught our AI coding agents doing something frustrating. Our prod deploy process was broken, and instead of flagging it, they just worked around it. Cursor found a workaround. Grok found another one. Both "solved" the immediate error by sidestepping the actual problem.
It took me longer than I'd like to admit to figure out why things kept feeling off. The deploys technically worked, but the underlying issue stayed broken—waiting to bite us at the worst moment.
The fix wasn't a better prompt or a smarter model. It was setting a clear expectation: when something is broken, stop and solve it. Don't improvise. Don't be clever. Flag the problem and either fix it properly or let me know so we can decide together.
Now my agents have a simple instruction: if a step in a process fails, don't skip it or patch around it. Stop. Tell me what's actually wrong. We fix the root cause or we document why we're accepting the workaround.
The clever path isn't always the right one—especially when "clever" just means hiding a problem for later.
Top comments (0)