Claude Platform adds OpenTelemetry tracing for agents in public beta

What did Anthropic launch for agent tracing?

On 14 October 2026 Anthropic announced OpenTelemetry traces on the Claude Platform, in public beta. Claude Managed Agents, the Messages API and the Claude SDKs can now send traces to any tool that accepts OpenTelemetry, so each agent turn, model call and tool use shows up as a standard trace with latency, tokens and IDs. Teams that already export Claude Code or Cowork telemetry can see the agents they build on the platform in the same place.

OpenTelemetry is the vendor-neutral standard most observability platforms speak. A trace is the full record of one request through a system; a span is one step inside it, such as a model request or a tool call. Anthropic says its spans follow the OpenTelemetry GenAI semantic conventions, so platforms that understand those conventions can display Claude data without extra setup and next to data from other model providers.

Anthropic frames the feature around trust. Its announcement argues that giving an agent more access should start with tracing its model calls, tool use and MCP server calls, reviewing those traces, and widening access as the agent proves reliable. The documentation adds more practical uses: debugging failed or slow requests, tracking token spend and latency, and building evaluations from real traffic.

How do API tracing and SDK tracing differ?

There are two ways in, and the documentation recommends using both. API tracing is configured once per organisation: an admin adds a destination in the Claude Console, and Anthropic’s servers export spans for every client, with no code changes. SDK tracing runs in your application: you install a recent Claude SDK and register an OpenTelemetry tracer provider, and the SDK emits a span for each request it sends and each tool its tool runner executes.

The split matters because each side sees different work. Only the API can see each model run inside a request, Managed Agents turns and the server tools Anthropic runs. Only the SDK can see client tools and tools in a self-hosted sandbox. Used together, the SDK sends W3C trace context with each request, and Anthropic’s server spans nest under your application’s own trace, linked to the user or feature that made the call.

For Managed Agents, which AI Lookout covered when Anthropic added dynamic workflows, a session gets a span for each Sessions API request and a turn span for each agent turn, containing model requests, tool uses and outcome evaluations. Spans for one thread can be grouped by gen_ai.conversation.id and a whole multi-thread session by anthropic.session.id. The SDK guide lists Python, TypeScript, C#, Go, Java, PHP and Ruby. One exception: for Claude Platform on AWS, the API spans arrive in their own traces rather than nested under yours.

What do the traces contain, and what stays private?

Every destination and every SDK client has a content mode. Metadata only is the default and exports identifiers, timing, status, token counts and text you choose to name, such as agent, tool and MCP server names. Full content adds the system prompt, input messages, model output and tool arguments and results. Anthropic notes that this content is then stored under your observability platform’s access and retention rules, so those settings become part of your data policy.

Some organisations cannot turn the content on at the platform level. If an organisation has a HIPAA, FERPA or K-12 data-handling restriction with Anthropic, Console destinations export metadata only whatever mode is selected, although SDK spans emitted by the customer’s own code are not affected. Zero data retention customers can still use Full content for Messages API requests, with the exported content held only in memory until delivery. Managed Agents is not eligible for zero data retention, and tracing does not change that.

There are size limits too. With Full content, each span carries up to 0.75 MB by default, which Anthropic says fits a 200,000-token context of English text and stays under the 1 MB request limit of many observability back ends. A full one-million-token context needs about 4 MB. Over the limit, spans are truncated and flagged, with system prompts and the earliest messages dropped first.

Which tools can receive Claude traces?

Any platform that accepts a standard OTLP endpoint works, and Anthropic names 12 partners: Arize AI, Braintrust, Databricks, Datadog, Grafana Cloud, Honeycomb, Langfuse, LangSmith, New Relic, PostHog, Sentry and Splunk. Its announcement groups them by job: seeing agent traces beside logs and metrics in tools like Grafana Cloud, Honeycomb, Sentry and Splunk; turning traces into evaluations in tools like Arize, Braintrust, Langfuse and LangSmith; and tracking token use across teams through Databricks.

The announcement quotes early customers. Ramp’s David Traina says the company’s Reflect system, which scores Langfuse traces and proposes fixes, has already cut token usage by 10% to 20% and made sessions 30% faster on some agents; that is Ramp’s figure for its own system, not a measured effect of the new export. Foxglove says it analyses agent interactions in PostHog next to product usage data, and Sterling Labs says it streams traces into Langfuse without custom instrumentation.

The table below shows why the server side matters. Client-side instrumentation can only see the requests an application sends, while work Anthropic runs on its own servers, such as Managed Agents turns, each model run and server tools, appears only in the API spans configured in the Console.

How should teams count tokens and costs from spans?

Token totals are easy to double count, and the documentation spends a full section on it. Usage appears on several spans of the same request, so you should sum one source per request. When a request has both an SDK span and an API span, both named anthropic.messages.create, Anthropic says to sum the API spans and tell them apart by span kind, client versus server. For Managed Agents, sum the model request and outcome evaluation spans. Compaction tokens need separate handling, and cache reads are already inside the input token count.

Spans can also arrive twice when a delivery is retried, so a platform that does not remove duplicates should count each span once by trace and span ID. More importantly, Anthropic says exported spans are not a complete record of API usage. Delivery is best effort, and a caller can drop a request’s spans by switching off the sampled flag in its traceparent header. To track each API key, for example to spot a leaked key, the documentation points to the Usage API instead.

What are the limits of the beta?

The feature is a beta under the Beta Services section of Anthropic’s terms. The SDK guide warns that SDK options and the span schema can change in ways that break code, queries or dashboards, for example by renaming a span, attribute, option or environment variable, and tells teams to pin the SDK version. Anyone building alerts on span names should expect to revisit them.

Traces also show what an agent did, not whether it should have done it. They are an input to the kind of review AI Lookout described in its guide to building an agent evaluation suite, and to the insider-risk framing Microsoft’s Satya Nadella set out in Models as Insider Risks. A sensible rollout starts in Metadata only mode, confirms that turns and tool calls appear where expected, and switches on Full content only for the workloads and retention settings that justify it.