AI Can Now Execute the Wrong Idea for Much Longer
Frontier agents can run for most of a working day, and the industry's favourite chart cheers the endurance. The clock that starts when you ask and stops when you accept is timed by almost no one.
AI progress has moved into the time dimension: frontier agents now run for hours, and the industry's favourite benchmark measures exactly that: how long they can keep going. Direction is taken on faith. A long-running agent carries its first interpretation of your request all the way through, so a minute-one misunderstanding compounds into polished, tested, wrong architecture. The clock that matters, from expressing an objective to accepting the result, is measured by almost nobody, and an agent can improve on every published benchmark while making it worse. As execution becomes abundant, everyone with an agent becomes a manager, and the scarce skills are the manager's: strategy, definition, and the well-timed interruption, because every ambiguity left in a brief will now be executed thoroughly, competently, and at length.
