OnTrack: Real-Time Agent Monitoring via Streaming Optimal Transport

OnTrack is a monitoring system that compares an autonomous agent's unfolding trajectory against reference runs in real time using streaming optimal transport. It flags divergence in about 1 ms per step, and its abort policy saves roughly 18% of compute on failing runs.

• Trajectory comparison latency is about 1 ms per step
• 83% precision: 5 of 6 aborted runs were truly failing
• Abort policy saves roughly 18% of compute on failing runs
• Catches loops, stalls, repeated tool calls and plan violations