Snowflake's Agent Observability hits private preview

AI agents produce more telemetry than traditional application requests because a single interaction can involve multiple prompts, tool calls, retrievals, retries and completions. That structure makes it hard to pinpoint where quality degraded or what drove token consumption, and retaining all the data can push observability costs up sharply. Snowflake's answer is Agent Observability in Observe, which the company says is built to retain and query high-fidelity agent telemetry at scale. It will offer an optional OpenTelemetry-compliant SDK, an Agent Explorer for searching traces and conversations, online LLM-as-judge evaluations for production traffic, and APIs for querying telemetry programmatically.

The product also aims to connect agent performance and AI spend to business outcomes. Snowflake says token totals alone do not show whether spend is producing value, so Observe derives latency, error, token usage and estimated cost signals from traces, letting teams compare agents, models and workflows. Combined with Snowflake's business data, teams could correlate agent behavior with customer experience, resolution rates, conversion and revenue, or with internal task completion and employee productivity.

Snowflake is emphasizing open standards as a hedge against lock-in. Observe supports OTel-native collection through an OTLP endpoint, and its storage commitment extends to Apache Iceberg. Whether that flexibility, combined with the promised cost economics of the Telemetry Lakehouse Foundation, is enough to win teams already juggling fragmented telemetry across frameworks and cloud services is the open question. For now, Agent Observability is a private preview, and its real-world performance at scale remains to be demonstrated.