Claude agent traces in your own monitoring: what the OpenTelemetry beta changes

On 14 October 2026 Anthropic announced OpenTelemetry trace export for the Claude Platform, in public beta. It covers Claude Managed Agents, the Messages API and the Claude SDKs. Traces record agent turns, model calls and tool use, with latency, token counts and IDs, and follow the OpenTelemetry GenAI semantic conventions, so any tool that accepts a standard OTLP endpoint can receive them.

Anthropic names twelve platforms it works with: Arize AI, Braintrust, Databricks, Datadog, Grafana Cloud, Honeycomb, Langfuse, LangSmith, New Relic, PostHog, Sentry and Splunk. API tracing is set up as a destination in the Claude Console, which needs the admin role, and works on the Claude API and Claude Platform on AWS. SDK tracing is set up in code and needs no role.

Each destination has a content mode. Metadata only is the default and holds identifiers, timing, status, token counts and names you choose, such as agent, tool and MCP server names. Full content adds the model prompts and responses. Anthropic's documentation says that after delivery the data is held by the service you chose, outside Anthropic, under that service's retention and access controls.

Organisations with a HIPAA, FERPA or K-12 data-handling restriction with Anthropic get metadata only from API destinations whatever the setting; SDK spans are not affected. Full content can be used with zero data retention on the Messages API, where the content is held only in memory until it is delivered. Claude Managed Agents is not eligible for zero data retention, and its retention does not change when tracing is on.

The export puts each model call, tool use and agent turn in the same monitoring stack as the rest of the estate, which is where operations and security teams already look. For a team running agents in production, that is likely the quickest route to answering why a run failed, what it cost and which tool it called.

Full content is a data protection decision, not a monitoring setting. Prompts and responses can hold customer records, contract text or staff data, and switching it on copies all of that into a second system with its own retention, access list and location. A data protection officer will likely want that second system on the record of processing and covered by its own contract before the setting changes.

The workspace setting gives a way to split the risk. A destination can cover all workspaces, selected ones or all except some, so a team can send full content from a test workspace and metadata only from the workspace that handles live customer data.

Anthropic's documentation says exported spans are not guaranteed to be a complete record of API usage. Delivery is at least once, so a span can arrive twice, and not guaranteed, so a span can be lost and is not sent again. A caller can also leave a request out by turning off the sampled flag in its traceparent header. To track each API key's usage, for example to spot a leaked key, Anthropic points to its Usage API, which counts requests without spans too.

The API retries a limited number of times after timeouts, connection failures and 408, 429 or 5xx responses. It disables a destination that sends a redirect, resolves to a non-public IP address, or fails every delivery for 24 hours, and sends nothing to it until someone enables it again. An organisation can have up to 10 destinations, including disabled ones.

This makes the traces good for operations and weak as evidence. They are the right tool for debugging slow or failed runs, tracking token spend by team and building evaluation sets from real traffic. They are the wrong tool to show an auditor or regulator that every agent action was recorded, because the documentation says plainly that some will not be.

The sampled flag matters for control. Whoever writes the calling code decides whether a request is traced, so tracing alone cannot prove that a team followed a monitoring policy. Spend reporting and key monitoring belong on the Usage API, with traces used to explain the numbers rather than to produce them.

A disabled destination fails quietly from the agent's point of view: the agents keep running and the traces stop. Someone has to own the delivery status, or a monitoring outage on a Friday can mean a weekend with no traces.