Claude Agent Traces Now Show Up in Langfuse

Debugging a Claude agent by reading logs and guessing is a rough way to spend an afternoon. The Claude Platform now sends its OpenTelemetry traces straight into Langfuse, and the feature just entered public beta.

That means the run you are staring at comes with its turns, tool calls, timing and cost already attached.

What Happened

The Claude Platform exports OpenTelemetry traces natively from three places: Claude Managed Agents, the Messages API and the Claude SDKs. The export is in public beta.

Once those traces land, Langfuse shows you the parts of a run that usually stay hidden. You get each turn, every tool the agent reached for, how long each piece took, and what the whole thing cost.

Why It Matters

Latency and spend are the two things that bite hardest when an agent moves from a demo to real traffic, and both are now sitting in the same view as the reasoning that produced them. You can see which tool call stalled instead of inferring it from a timeout.

The other half is evals. Langfuse lets you turn real runs into evals, so the failures your agent actually hit become the test cases you improve against, rather than scenarios you made up in advance.

If you already have agents running on the Claude Platform, the tracing is worth switching on before your next debugging session.