Tessary

Tessary traces

A trace is one unit in Tessary: the stored record of one agent turn. Traces live in Tessary's substrate, the agent-native store every other part of the product reads from. The substrate has three fixed levels: the session, the trace (one turn), and the span (one step, an LLM call or a tool call). The levels match how agents actually run: a user session holds turns, and each turn holds the steps the agent took to produce its answer.

Each level is keyed on your own ids, and a span carries the actual content of its step: the payloads, tool calls, retrieved documents, and media. Spans can also carry tessary.call_site.id, a plain span attribute that needs no SDK. It's what findings, baselines, and regressions bind to.

To send traces, point the OpenTelemetry exporter your agent already uses at Tessary over OTLP, or use the SDK push source if your stack doesn't have one. PII redaction runs on the write path, before anything is stored. Classifiers, cases, and root cause analysis all read this same stored record, so every conclusion cites evidence you can open.

6 questions

Answered, plainly.

Does Tessary read my Langfuse account?No. Tessary has no Langfuse connector and never pulls from an account. You add it as a second OpenTelemetry exporter pointed at Tessary's OTLP endpoint.answer →Does Tessary store images from my agent's traces?Yes. Images in a span's input or output are stored and deduplicated, so the same screenshot reused across traces is kept once, not copied per trace.answer →Is Tessary a trace explorer?Not primarily. Tessary stores every trace as evidence, but the viewer is built for confirming a case a classifier already flagged, not open-ended browsing.answer →What does a production span carry that a dataset row cannot?Fields like session.id, deployment.environment.name, and the tool call that actually ran, context that only exists because a real request happened.answer →What is a trace?A trace is Tessary's stored record of one agent turn: one request a user waited on, and the spans that ran to produce its answer.answer →What structure does an ingested agent trace have?Three fixed levels: a session holding a user's turns, a trace for one turn, and the spans inside it, each an LLM call, tool call, retrieval, or sub-agent step.answer →

Send us the traces you already emit.

book a demo