Telemetry monitor lifts agent success to 73%

A lightweight telemetry monitor plus rollback and retry raised agent task success from 52% to 73% across 2,823 episodes in the Agent Trajectory Sentinel study, while running at roughly 200 microseconds per step. Deterministic checks caught 96% of failures when call coverage was included with zero false positives in the 63-case comparison, but the learned monitor needed per-deployment recalibration and recovered only 45% of flagged failures. Teams building the kind of containerized agent regression loop Supabase demonstrated should log tool telemetry and verify outputs before paying a second model to judge every step.