Researchers have proposed a new method for monitoring the behavior of long-horizon agents, which are AI systems that operate across multiple steps and tasks. The approach, called ontological trust, estimates whether the agent's trajectory is still aligned with its original task and user authorization. This is particularly important as agents can drift away from their intended goals over time, even if they make individual decisions that seem valid locally. The researchers introduce a new online monitor called RGE, which uses large language models to derive structured representations of tasks and steps, but makes deterministic updates to trust states. They demonstrate the effectiveness of RGE on a cross-domain corpus of trajectories, outperforming existing methods in detecting drift and maintaining high accuracy while keeping false positives low. --- Why it matters: This matters because long-horizon agents are increasingly used in real-world applications, such as finance and healthcare, where their behavior can have significant consequences. The ability to monitor and control these agents' trajectory is crucial for ensuring that they remain aligned with their intended goals and user authorization. Source: https://arxiv.org/abs/2608.17718