Tessary

Tessary is an open-source reliability platform for AI agents running in production, developed by Tessary AI. It monitors every trace, flags problems with classifiers, groups related findings into cases, and runs root-cause analysis against trace and repository evidence. It is available as Tessary Cloud or as a self-hosted deployment under the Apache 2.0 license.

What It Is

Tessary targets engineers whose agents complete runs without errors but produce wrong results. Agents send OpenTelemetry traces to Tessary, which reads all of them rather than a sample. Classifiers check each trace cheaply, and the output is a case with a summary, the likely causes, and the traces behind them.

How the pipeline works

Traces arrive over OTLP (HTTP and gRPC) or SDK push and are normalized to the OpenTelemetry gen_ai conventions, with PII redaction before storage. Built-in classifiers cover behavior drift, tool errors, secret leaks, malformed output, frustration, groundedness, and cost and duration drift. An LLM triage step decides whether a finding is a real deviation before it becomes a case. RCA then runs an agentic session over the failing traces and, if a GitHub repository is connected, the code itself. It returns the causes found and the checks it ruled out. Alerts route the case to its owner; Tessary explains and hands off but does not open the fix.

Setup path

Tessary Cloud can be used by signing up. Self-hosting uses a Docker Compose command or a setup guide written for a coding agent, and requires Docker Engine 26+ and Docker Compose v2.34+. To instrument an agent, point OTLP traces at the endpoint and add a tessary.call_site.id attribute to spans that invoke a model.

Self-hosting and telemetry

Self-hosted instances send an anonymous heartbeat to the vendor, which can be disabled with an environment variable. The README states it never carries trace or prompt content.