OpenTelemetry for the Claude Platform

Anthropic just shipped OpenTelemetry for the Claude Platform in public beta, and it fills an important gap in agent observability.

Our own services were mostly traced, but the provider side wasn’t. One Messages call that ran three model requests and a server tool showed up as a single span with a duration and a token count. Now that call decomposes into model request, tool use, and compaction spans from inside the API.

It’s an even bigger deal for Managed Agents, where the entire loop that used to run out of sight is now traced turn by turn.

Setup is easy with no application code change and the spans follow GenAI semantic conventions so any OTLP collector take them as-is.

By Anthropic’s own account, delivery is best effort, so we need to treat it as debugging visibility and keep our audit trail on the Usage API.

Now we can start the agent with narrow access, trace everything, and widen permissions only as the traces show it behaving.