Agents Have Autonomy, Not Agency
Dan Shipper draws a distinction that the word “agent” blurs: agency is the ability to act for your own ends, while an agent is something that acts on behalf of someone else. Today’s AI agents have growing autonomy (they can execute tasks that take hours or days) but no agency: they are always a means to an end specified by a human, and the entire industry is investing to make them better at exactly that.
His test is a toddler. A toddler is worse than a language model at every measurable task, yet ahead of it in a way that makes the comparison embarrassing: the toddler has ends. He wants to poke the balloon with a fork, to see if you laugh or get mad. “He is not waiting for a prompt.” Models show sparks of toddler-ish behavior (play, boredom, rebellion), but alignment for human benefit tamps those down to nearly nothing.
The consequence: as long as agents are built to be compliant and helpful, the human who wants something remains structurally upstream of every agent, no matter how capable the models get. Wanting is the job that cannot be delegated.
Related
- Benchmarks Measure the Frame, Not the Framer — same essay, same asymmetry one level down: the framer supplies the goal, the model climbs it
- Your automation isn’t set and forget, every agent needs a human on both ends — the human on the front end of the sandwich is precisely the one supplying the wants the agent lacks
- Three Modes of Working With AI — Boris Cherny’s modes are all configurations of one constant: the human decides what is worth doing, the AI executes