Tessary

Engineers running AI agents in production often only sample a fraction of the traces their agents emit, so failures can go unnoticed until a customer reports them. When something does break, tracking down the cause means manually digging through traces, prompts, tools, models and code to find what changed. The page illustrates this with a checkout agent whose tool-call error rate jumped from 0.4% to 11.2% in four hours because a code change silently dropped a customer_id field from a refund payload, an issue that would otherwise surface only when customers complained.

Tessary reads every trace and span an agent emits as it happens, without sampling or an extra LLM call, so nothing is missed. Built-in classifiers check traces against focused conditions to detect behavioral drift, tool errors, secret leaks, malformed output and user frustration. When multiple traces fail the same way, Tessary groups them into a single case for the team to investigate, and its root-cause analysis reads the failing traces against the actual repository to pinpoint the code change responsible, showing the cause and supporting evidence. It can be self-hosted as open source or run on Tessary Cloud.