Research: Agent Behavior Shaped More by Harness Than Underlying LLM

dair_ai · x · 2026-08-28

New work collapses agent traces into finite-state machines (FSM), showing complex agent behaviors can be represented with just 7-43 states. FSM context outperforms agent workflow memory in next-step prediction and failure prediction (AUROC 0.94). The authors suggest behavioral topology is shaped more by the deployment harness than the LLM itself.

Original post →

More from coding & agent

coding & agent channel →