Different LLMs may behave more similarly inside an agent harness than their raw traces suggest.

LLM agents can produce massive traces while following a surprisingly small behavioral graph, and that graph can predict both what happens next and when a run is failing.

Basically this paper finds that AI agent runs can be reduced to a small action-state map that predicts next moves and failures, so teams should monitor structure, not just raw traces.

It turns many past agent runs into 1 compact finite-state machine: a map of recurring actions such as search, edit, execute, and submit.

Across 12 datasets, these maps needed only 7–43 states and could replay held-out traces with at least 0.997 fitness.