the log
What we measured.
latest ·
Describe a failure only your agent has, and Tessary builds a small judge to check every session.
turning intent into a judge for your agent
Tessary's custom classifiers turn a failure only your agent has into a small judge, built with you from your real traces, that checks every production session.
read the post →newest first
5 posts
a browser tool looked 27x faster because it had started failing
p50 latency fell from 30.1s to 1.1s while the error rate rose from 2.4% to 6.2%.
finding RAG answers that aren't grounded in their sources
Our groundedness model caught 39% of unsupported RAG answers at a 2% false-alarm rate, with 83% precision.
how we built a frustration classifier for every conversation
Jev matched Sonnet 5's recall on flagging unhappy users at roughly 1/34th the cost.
a shared tool started failing on 1 in 7 calls
Failure rate for a shared write tool climbed from 1.7% to 14.8% over a week.
AI agent failures we found across 22 companies
Most of them returned a normal-looking response and passed sampled evals.
Try Tessary on Cloud.
free · 10,000 traces a month · 1 GB · 30-day retention · $10 one-time credit for triage and RCA · no credit card