Models don't have agency, systems do: the one-token trick that seeds tool use

sethjuarez · x · 2026-09-16

In an 18-minute essay, sethjuarez argues that a language model only ever emits tokens — everything we call "agentic" is a story about shaping those tokens into intent a runtime can act on. The model never gains agency; the system built around it does.

Ask a model to book a flight and it replies with a friendly sentence a human can parse but a program chokes on — the words are there, the shape is not. Closing that gap is what hands a system agency.
Using the philosophy-of-action definition (an agent has the capacity to act from intention), a language model fails outright: it originates no intent and performs no action.
The practical seed: constrain the model to answer BOOK or CANCEL, nothing else, so if reply == "BOOK" just works — a paragraph of intent becomes one token your code can branch on.
The essay traces an orderly path from a word, to a structure, to a function: making the model useful to software by shaping what it may say, not by making it smarter.

Original post →

Models don't have agency, systems do: the one-token trick that seeds tool use

More from coding & agent