Agent observability
Learn how you can view and use telemetry data from your AI agents.
Private preview feature (Invite only)This feature is enabled per tenant by Observe by Snowflake and is not generally available. To participate, contact your Observe by Snowflake representative. You may access and use this preview feature solely for internal evaluation purposes. Previews are provided as-is. Your use of this preview is subject to the Snowflake Preview Terms located at https://www.snowflake.com/en/legal.
Agent observability allows you to monitor, debug, and score your AI agents in Observe by Snowflake. Traces show what the agent did. Evaluations (evals) score whether the answer was any good.
With an instrumented agent you can perform tasks such as:
- Search conversations in AI Agents. Each turn is a trace. Model calls, tool calls, and retrievals are spans on that trace.
- Read latency, errors, token usage, and cost from metrics built off those spans.
- Run online evals on production traffic and attach the score to the spans the judge saw.
- Send results from an offline eval run into Observe as traces or metrics.
- Query traces, metrics, and eval scores from the API.
The JavaScript and Python SDKs follow the OpenTelemetry GenAI semantic conventions. They auto-instrument the OpenAI SDK, the OpenAI Agents SDK, LangChain, LangGraph, and the Claude Agent SDK. The Vercel AI SDK and the Anthropic SDK are JavaScript only. Any other stack can emit the same spans with the Client API. See the framework table in Instrument your agents.
To learn more, check out our blog post: Improve AI Performance, Quality and Cost with Agent Observability.
Updated 41 minutes ago