One sitting versus many separate sittings
Think about the difference between answering one trivia question and coaching someone through training for a marathon. The trivia question happens once, start to finish, in a single conversation. You hear it, you think, you answer, you're done. Coaching a marathon runner is nothing like that. It happens in pieces spread across months: a workout on Monday, a rest day, a check-in on Friday, an adjustment after a bad week, and so on. Nobody re-explains the entire marathon goal at the start of every single workout. The coach has to remember where the runner left off and keep steering toward the same finish line, session after session, without a constant refresher.
That's the real line between a normal agent task and a long-horizon one. A normal, single-turn agent task is like the trivia question: the agent gets a request, works on it inside one continuous stretch of context, produces a result, and its job is over. It only ever has to be right once, in one unbroken train of thought. A long-horizon task is like coaching the marathon: it's a big goal that has to be pursued across many separate sessions, possibly by a version of the agent that has never seen the earlier sessions at all, and it still has to end up in the right place.