A new wave of agents shipped this week promising to work for hours without a human in the loop. Give it a goal before coffee, come back to finished output by lunch. The pitch is the same one autonomy always sells: hand off a whole day, not just a task.
I understand the appeal. I also think everyone is measuring the wrong number. Duration is the headline. Duration is not the feature. It's just exposure.
Here's the mechanism. An agent that runs for five minutes gets five minutes to go sideways before anyone notices. An agent that runs for five hours gets five hours. The failure surface doesn't shrink because the leash got longer. It grows, quietly, in a room nobody's watching. A wrong assumption made in minute six doesn't stay small. It compounds through minute two hundred, and by the time you check in, it isn't a typo anymore. It's the whole afternoon.
So the real question was never how long an agent can work. It's how easily you can see what it did, and how easily you can undo it. That's a checkpoint. Not a progress bar — an actual pause, mid-run, where the agent shows its work so far, waits for a look, and can be rolled back cleanly if the look says no. A system with checkpoints and a four-hour runtime is safer than a system with no checkpoints and a twenty-minute one, because the twenty minutes still runs unsupervised in the dark.
Think about how you'd actually manage a person doing a long, open-ended job. You don't hand someone a brief at nine and disappear until five. You check in at eleven. You look at the draft before it becomes the deliverable. Not because you don't trust them. Because that's how anyone catches a wrong turn while it's still cheap to fix. Long-running agents deserve the same courtesy we'd give a long-running colleague, and most of the products racing to ship "hours of autonomous work" are quietly skipping that part, because a pause button is a harder feature to build than a longer timer.
This matters for anyone actually shopping for these tools, not just admiring the demo. When you're evaluating a long-running agent, don't lead with how many hours it can go. Lead with this: can I see its state at any point mid-run. Can I interrupt it without losing the work it already got right. Can I roll back the last step, not just kill the whole job. If the answer is no to any of those, you haven't bought autonomy. You've bought a longer blind spot.
None of this is an argument against long-running agents. It's an argument for building them like anything else that operates while you're not looking — with brakes, not just an engine. The teams who get real leverage out of this next wave won't be the ones who let their agents run the longest. They'll be the ones who can see inside the run at any point and trust what they see enough to let it keep going.
Takeaway: before you ask how long an agent can run unsupervised, ask how easily you can see inside the run and stop it. Runtime is the headline; the checkpoint is the feature. ✱
