As of October 14 you can send OpenTelemetry traces from Claude Managed Agents, the Messages API and the Claude SDKs to any OTel tool. It’s a public beta. If you already pipe Claude Code or Cowork traces into your observability stack, the agents you build yourself now show up in the same place.
What the traces contain
You add a destination in the Claude Console, and from then on every agent turn, model call and tool use arrives as a standard trace, with latency, token counts and IDs. Managed Agents and the Messages API need that destination; the SDKs export on their own. The SDK side covers Python, TypeScript, C#, Go, Java, PHP and Ruby.
The semantics matter more than the export
Claude’s traces follow OpenTelemetry’s GenAI semantic conventions. That sounds like small print, but it decides whether the data is usable at all: it’s the only way Claude and other providers land in one dashboard under the same field names. Claude Code 2.1.300 picked up the same attributes on its trace spans, so the two fit together.
Named in the announcement: Grafana Cloud, Honeycomb, Sentry and Splunk for debugging next to logs and metrics, plus Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic and PostHog for turning real traces into evals. Databricks Unity Gateway is aimed at platform and security teams who want token usage broken out by team. Any other OTLP endpoint works too.
What customers built with it
Ramp runs an agentic system called Reflect that scores its own Langfuse traces, finds recurring failures and proposes fixes. On some agents that has already cut token usage by 10 to 20 percent and made sessions 30 percent faster. At a finance customer, every model call sits in the same trace as the journal entry it ran inside, so when an auditor asks what the agent did, the answer is a timeline rather than a story.
Watching comes before permissions
The most interesting line in the announcement sits right at the top and reads like an operating rule: trusting an agent with more access starts with tracing its model calls, tool use and MCP server calls. Review the traces, then widen the access. In practice that order is almost always reversed, and it is at least now technically available without writing your own instrumentation.