Researcher breaks down agents: the model just predicts, the surrounding machinery turns predictions into actions

vishalmisra · x · 2026-09-27

Researcher Vishal Misra lays out a structural view of what makes agents different from plain LLMs:

The difference is the machinery around the model, not the model itself: the model still just does next-token prediction. Tools, memory, permissions, feedback, and repeated calls are what turn predictions into actions, state, and loops.
Three research threads converge: he notes that an earlier talk, the recent Bayesian interpretation of attention, and the "show me the loop" work all fit together, answering three questions: what structure does the model learn, how does it update, and what machinery turns predictions into actions?

A concise mental model for agent architecture: model capability stays constant; the engineering scaffolding defines agent behavior.

More from coding & agent