What Agents Can and Can't Do

Something can be capable and still not be trustworthy, and the gap between those two words is where most of the disappointment lives. A system that does a hard job brilliantly nine times out of ten is not ninety percent of a useful system. It depends entirely on what the tenth time costs and whether anyone notices.

That turns out to be the question that decides where this technology works, and it has less to do with difficulty than people expect. Some hard tasks suit it. Some easy ones are a disaster. What separates them is whether the world answers back: whether a wrong step produces an error, a failed test, a visible mess, something that says no before the damage is done. Work that fails silently is where things go wrong, no matter how simple it looked.

Underneath sits a harder problem that does not get solved by better engineering. You have to say what you want, and what you say is never quite what you mean. Every unstated assumption, every obvious exception you did not think to mention, is a gap something can walk straight through while following your instructions perfectly.

That gap has a name and a long history, and nobody has closed it.