New Monitor Assesses Trust in Long-Horizon AI Agent Trajectories

As AI agents become more complex and operate over extended periods, using multiple tools and observations, simply verifying individual actions is insufficient. The critical challenge is ensuring the agent's overall trajectory remains aligned with the user's original intent, preventing subtle "drift" where the agent might perform plausible actions but deviate from the authorized task or objective. Current monitoring tools often focus on local compliance or generic risk, failing to assess this prefix-level alignment. This research proposes "ontological trust," a task-conditioned property for agent trajectory prefixes, and introduces RGE (Role, Goal, Evidence) as an online monitor. RGE uses Large Language Models (LLMs) to create structured representations of tasks and steps, but its trust-state updates and intervention decisions are deterministic, ensuring auditable and replayable trust trajectories. Evaluated across diverse domains, RGE significantly outperforms existing baselines in detecting prefix-paired drift, achieving high F1 scores while maintaining good benign coverage, though detecting "pseudo-consistency" failures remains challenging depending on external visibility of task completion.