LLM agents are increasingly used in settings where mistakes are expensive, such as trip planning, stock trading, and IT incident triage. These agents often operate with little rule-based oversight, and some actions cannot be undone once taken. That combination creates a need for monitoring that can catch problems while there is still time to intervene.
OnTrack, described in a new arXiv preprint, approaches this by treating an agent's trajectory as a structured stream of states and using optimal transport to compare it against expected or safe paths. The method is designed to be real-time and structure-aware, meaning it can flag deviations as they happen rather than after the fact.
By intervening during execution, the system aims to reduce both cost and safety risks associated with autonomous agent behaviour. The paper does not report experimental results in the abstract, so the practical effectiveness of the approach remains to be seen.