Snowflake introduces AI Agent Observability for LLM application monitoring

This new capability enables teams to monitor, debug, and improve AI agents and LLM applications by tracing interactions, connecting agent behavior with cost and quality, and retaining telemetry cost-effectively at scale. It tackles new concerns around response quality, AI spend, and the complexity of multi-step agent workflows.

Teams can instrument agents using an OpenTelemetry-compliant SDK with support for LangChain and OpenAI/Anthropic Agents SDKs, or send custom traces via an OTLP endpoint. Agent Explorer allows searching and debugging conversations, while metrics derived from traces help monitor performance, usage, and cost. It also supports online LLM-as-judge evaluations for production traffic to surface quality issues and integrates offline evaluation results.

Agent Observability provides agent-aware views across conversations and traces, helping engineers inspect prompts, completions, and tool activity to pinpoint failure origins. The Context Graph connects agent telemetry with related applications, services, and infrastructure for end-to-end debugging. It supports online LLM-as-judge evaluations for production quality signals and connects results to underlying traces to assess changes.

Observe is designed to retain and query long-running, context-rich agent traces and conversations at scale, addressing the challenge of high telemetry volumes from agent interactions. It leverages cloud object storage and a Telemetry Lakehouse Foundation, separating storage from compute to retain more data and scale compute independently for cost efficiency.