AI Agents Speak OpenTelemetry: Two Signals from MCP and OTel

On July 28, 2026, Model Context Protocol released a new specification that officially adopts W3C Trace Context for passing trace context across MCP journeys. The spec defines three keys in the _meta field of params: traceparent, tracestate, and baggage, with an exception to the DNS prefix rule to ensure interoperability. This practice was already working in SDKs and tools like Logfire, OpenInference, and Envoy AI Gateway. Additionally, MCP deprecated built-in Logging, Roots, and Sampling, recommending OpenTelemetry for structured telemetry, with a transition window of at least twelve months. The core protocol became stateless, removing initialize, initialized, and Mcp-Session-Id, with version and trace context now sent in _meta of each request. On the same day, Japan Community Day and KubeCon Japan featured talks on AI agent observability with OpenTelemetry, which after graduating in May 2026 listed priorities including semantic conventions for generative AI and agentic scenarios, browser and mobile observability, schema management via Weaver, simplified installation via Packaging, and the Injector library. The articles emphasize that one trace_id is insufficient for AI systems, and semantic conventions are still evolving. The news concludes that AI agents must be operated as distributed applications, using W3C Trace Context, OpenTelemetry spans, and semantic conventions to collect data in-house, as with the Proto Observability Platform.