The First and Last Harness Is Us
The first and last harness
The first harness was us
Look closely at every layer of the harness and you find a person underneath it. Models learned to be helpful because people rated their answers. The onboarding binder is written by a person. The keycard is issued by a person. The approval step waits for a person. Even the goal, the definition of a job well done, starts as something a human wanted.
Strip away the software and that's what a harness really is: human intent, made durable. The tools and loops are just how we hand our judgment to something that can act faster than we can.
That's also why the hardest problem in AI is a harness problem. Today we steer models by checking their work. Aschenbrenner's point about superalignment is that this stops working once models are smarter than the people checking them. You can't grade an exam you can't read. The question of the next decade isn't whether humans stay in the harness. It's whether we can keep steering when we're no longer the smartest thing at the desk.
The belt becomes the vault
Superman doesn't need a utility belt. But the Justice League still keeps kryptonite in a vault. That's the arc of the harness. Today it exists to help agents do more. But the smarter and more capable the models become, the more it will exist to prevent them from doing things they shouldn't.
So the question was never whether smarter models will still need tools. It's who builds the desk, who owns what's on it, and who holds the key to the door.