Agent-state-trained 4B models reportedly make AI agents 81.1%–89.0% cheaper on tasks like their training data

A post describing a Princeton University–Google paper says its agent harness lets a smaller model choose tools only when its confidence clears a threshold fitted on the agent’s own runs. On tasks like their training data, 4B models trained on agent states reportedly made agents 81.1%–89.0% cheaper and up to 2.5 times faster, with success staying within three points. Giving the smaller model every tool choice without a threshold lowered success by 19.0 points.