Snowflake Cortex AI Gateway introduces governed end-to-end agent traces
Snowflake details governed agent tracing through Cortex AI Gateway, combining model and MCP tool activity in Snowflake with role-controlled access and optional payload capture.
The public-preview gateway records model requests and MCP tool calls as OpenTelemetry spans. Payload capture is optional, and full application traces require additional client instrumentation.
Snowflake detailed how Cortex AI Gateway introduces governed end-to-end agent traces on 19 October 2026, recording model and Model Context Protocol (MCP) tool activity as OpenTelemetry spans inside Snowflake. The public-preview service captures usage, timing, failures and policy outcomes for requests routed through it. Activity that bypasses the gateway is not automatically visible.
In its official engineering announcement, Snowflake described tracing as part of the gateway’s serving path, rather than a separate monitoring agent that developers must install. The gateway itself entered public preview in September; the October publication explains its tracing and access-control capabilities in greater detail.
Gateway spans track usage and tool outcomes
Snowflake says each model request passing through the gateway creates a span recording the model, caller, token usage, cache activity, latency and completion status. Supported MCP tool calls can be recorded alongside inference, including the tool identity and whether a call succeeded, failed or was denied by policy.
Administrators can inspect the information through Snowsight or query it using the AGENT_TRACE_TABLE table function. Records can also be joined to AI_GATEWAY_USAGE_HISTORY by request identifier to connect credit consumption with individual requests. These records support usage attribution and incident investigation, but do not by themselves show everything an agent did outside the gateway.
Prompt content remains an opt-in choice
Logging request metadata is distinct from storing message content. Snowflake says administrators can enable payload capture for prompts, responses, tool arguments and results, but those contents are not recorded by default.
Traces reside in customer-owned Snowflake event tables subject to role-based access control. Reading gateway traces requires MONITOR permission. Disabling payload capture stops future content collection; it does not delete payloads already stored.
Client instrumentation extends the trace
The gateway can adopt OpenTelemetry trace context from an instrumented client so its model and tool spans appear within the broader application trajectory. Additional client spans are necessary to inspect work the gateway cannot observe, such as local decision steps or retrieval operations.
Snowflake also supports bringing compatible external spans into the same trace view. It cites tools such as TruLens and OpenTelemetry-based instrumentation, but those integrations still require setup.
Public preview constrains availability
Snowflake’s current observability documentation identifies the gateway as an open-preview feature available across AWS, Microsoft Azure and Google Cloud commercial regions, with AWS Europe (Spain) excluded. Its 19 October release notes separately describe Cortex Tools for MCP connections as a preview capability.
The announcement does not establish general availability, universal cloud coverage or independent improvements in agent reliability. It introduces a governed way to investigate routed AI activity, rather than a complete record of every action an agent might take.