AI agent problems

AI agents mainly cause control problems: they can do things you didn't ask for, declare a task finished when it isn't or spend more than they should without you noticing. An agent is an AI that doesn't just answer but acts on its own on your files, your email or your tools, and that's exactly where the risk lies.

What usually confuses beginners is trusting the agent because it sounds so sure of itself: it saying "done" doesn't mean it's done well. The typical problems are broad permissions you grant without thinking, silent errors you don't see until days later, decisions it takes without consulting you, loops where it keeps repeating the same thing and costs that shoot up when you pay per use. None of this means they're useless; it means you have to supervise them like a very fast, very enthusiastic intern.

It's a topic close to my heart, because I work with agents, but there are no blog articles with this tag yet. In the meantime, my practical recommendation: give the agent the minimum permissions, have it work on copies rather than originals, check what it has done before accepting it and, if you pay per use, set a spending cap when the tool allows it. It's slower, yes, but undoing a disaster is much slower.