Claude Platform traces now export to LangSmith via OTel

Developers running agents on Anthropic's Claude Platform can now send execution traces to LangSmith, LangChain's tool for inspecting agent runs, without changing their agent code. According to the LangSmith documentation, the Claude Platform exports OpenTelemetry (OTel) traces of Claude Managed Agents sessions and Messages API requests straight from Anthropic's servers to a destination that an admin registers in the Claude Console. Adding LangSmith as that destination is a form in the console. The feature is in public beta.

That server-side path is one of three, and they differ in where the records come from. In the second, the Claude SDKs inside a developer's own agent emit an OTel span for each request they send and each tool use the tool runner runs, and an exporter library forwards those spans to a LangSmith endpoint. The third wraps the Anthropic client with LangSmith's own SDK, wrap_anthropic in Python or wrapAnthropic in TypeScript, which the documentation recommends for tracing Anthropic calls together with the rest of an agent's code in a single trace. Only the platform and Claude SDK routes are labeled public beta. The console route needs the admin role in the Claude Console organization, and a self-hosted LangSmith instance must be reachable over HTTPS on port 443 at a hostname with a public IP address. The platform disables a destination whose hostname resolves to a non-public address.

The setting with the most consequence is content mode. A destination set to full content includes prompts and responses in the spans, while metadata only omits input and output, and the Claude SDKs record metadata only unless configured otherwise. Full content is what lets someone read what the model was given and what it said when a run goes wrong. It also means that prompt and response text is stored in LangSmith, so a team turning it on is choosing to copy that text into a second service.