Models Don't Have Agency. Systems Do.: shaping tokens into actionable intent

sethjuarez · x · 2026-09-10

In an 18-minute essay, sethjuarez argues that by the philosophy-of-action definition (an agent has the capacity to act, driven by intention), a language model has no agency: it only emits tokens, originating neither intent nor action. The gap is that models produce sentences humans can act on but programs can't parse — "the words are all there, the shape is not." The fix isn't smarter models but constraining what they may say, shaping tokens from word to structure to function so a runtime can act. Agency belongs to the system built around the model, never the model itself.

Original post →

More from AGI Musings

AGI Musings channel →