Tessary

About

Tessary is an open-source agent reliability platform that monitors every production trace and investigates the issues it finds. Cheap L1 classifiers flag likely erroneous traces, then L2 agents run on SOTA models to triage only those flagged traces and find the root cause. The authors report that at 1M traces this was 5x cheaper than sampled evaluations, because expensive grading is not run across all traces.