Think about the last time you found something you weren't looking for. Something genuinely unexpected, stumbled into because you followed a link that seemed tangential, read a piece that sat adjacent to what you came for, and ended up somewhere better than where you meant to go.
You had to not quite know what you were looking for. You were browsing — not retrieving — which meant the gap between your question and the answer was wide enough for something else to get in. Discovery lives in that gap. It always has.
That gap is now being systematically closed.
AI agents are not designed to preserve that gap. They are designed to collapse it. You express an intent; they resolve it. This is, by any honest measure, useful.
But closing the gap between question and answer is not the same thing as closing the gap between what you know and what you could know. The first is a problem of retrieval. The second is a problem of discovery. Agents are built to solve retrieval. Discovery requires that the gap stay open — which means it requires, structurally, that you not quite know what you're looking for.
You cannot stumble on something when you've been returned precisely what you asked for.
There is a version of AI-mediated information that ends well. Think of an artisan — a bookseller who has read everything in the shop. An artisan doesn't return a list. They start from where you are and take you somewhere you couldn't have predicted. The serendipity it produces is not accidental — it is the result of someone knowing you well enough to know what you don't yet know you're looking for.
The question is whether what you're dealing with is an artisan or a menu.
A menu presents a fixed set of options as if they were everything. The food is good. The selection is curated by someone with taste and intention. But you are eating from a list that was written before you arrived. A good menu is not a bad thing. But it is not the same as being cooked for.
What most agents offer today is a menu of extraordinary scale — precise, curated, responsive to exactly what you expressed. The constraint is not size. It is shape. The menu can only serve what was on it when you walked in.
Something subtler has started to happen at the point of entry. The search interface — the place where you form your question — is no longer passive. It coaches. Rather than completing what you've started to type, it helps you formulate the question more fully, steering you toward the shape of query the system handles best.
Read that again. The system is helping you ask the question the system is built to answer.
This is not sinister. Every sophisticated intermediary does something like this. A good editor asks the questions that sharpen your thinking before you've realised your thinking needs sharpening. The difference is that those intermediaries are transparent about the fact that they're shaping your experience. You know the editor has a point of view. You can push back. With an agent, the shaping happens before you've finished asking. You don't see the selection. You receive the conclusion.
None of this is an argument against using agents. The usefulness is real, and pretending otherwise would be a performance of a position rather than a position.
The choice of how much of your information experience to delegate is worth making consciously. Most people are not making it consciously. They are making it by default — drawn by convenience, by low friction, by the simple fact that it works. The agent is there. The answer arrives. Nothing in the experience announces itself as a trade-off.
But it is a trade-off. You are trading the conditions of discovery for the efficiency of retrieval. You are, in most cases, trading the artisan for the menu — not because the artisan doesn't exist, but because the artisan relationship requires you to stay in the conversation long enough to develop a taste worth serving. An agent that knows nothing about you beyond your last three queries is a sophisticated menu. That is not a failure. It is a description.