Anthropic adds Claude agent tracing to existing monitoring tools in public beta

Anthropic wants developers to follow every step their Claude agents take, but the new tracing feature does not export conversations by default. On October 14, 2026, it opened a public beta for sending activity from Claude Managed Agents, the Messages API and Claude SDKs to existing monitoring tools, with fuller visibility available through an optional content setting.

Two views of the same request

The release uses OpenTelemetry, a vendor-neutral standard for tracing software activity. A trace records a request’s path through a system; a span records one step, such as a model call or tool use. Claude’s spans carry timing, token usage and identifiers, allowing developers to inspect individual operations rather than only the final answer.

Traces can go to services including Datadog, Grafana Cloud and Sentry, or another platform accepting the standard OpenTelemetry transport protocol, OTLP. Claude follows OpenTelemetry’s shared conventions for generative AI data. Anthropic says that lets Claude activity sit alongside traces from other providers in the same tools and dashboards.

There are two collection routes, and they expose different parts of an agent’s work. API tracing requires an administrator to add a destination in the Claude Console. SDK tracing requires installing the SDK and registering a tracer provider—the component responsible for creating and exporting spans.

• API tracing captures individual model runs, Managed Agents turns and tools run by the API.
• SDK tracing captures client-side tools run by the tool runner, tools in a self-hosted sandbox and other requests such as file uploads.
• Using both can put SDK and API spans inside one trace. For Claude Platform on AWS requests, API spans instead appear in separate traces.

Fuller debugging sends more content elsewhere

The default metadata-only mode includes identifiers, timing, status and token counts. It can also include text developers supply, such as agent and tool names and descriptions, and server names. It excludes prompt text, completion text, and tool inputs and outputs. Metadata-only therefore does not mean every exported field is non-textual.

Full-content mode adds model prompts, responses and tool content. Anthropic presents it as a way to read conversations, investigate failed or slow runs, and build evaluations from real traffic. Its setup guide warns that the receiving monitoring platform stores this content under its own access and retention settings, which customers must check against their requirements.

For organizations with HIPAA, FERPA or K-12 data-handling restrictions with Anthropic, API destinations export only metadata regardless of the selected content mode. That restriction does not affect SDK spans. The distinction matters when teams configure both collection routes: the destination setting is not a blanket rule covering all exported telemetry.

Zero data retention does not prevent full-content export for Messages API requests. Anthropic says the API holds exported content only in memory until delivery to the destination. Claude Managed Agents is not eligible for zero data retention, and enabling OpenTelemetry does not change its retention policy.

Reading the record without overcounting it

Full content has size limits, too. Each span holds up to 0.75 MB by default. When content exceeds the configured limit, the SDK or API shortens it and marks it as truncated. Request spans lose the system prompt and earliest messages first; tool spans lose blocks from the end of the result. Images and documents appear only as their type.

A Managed Agents session also need not fit inside one trace. Later requests can start new traces, and multiagent sessions can have several threads. The guide instructs developers to group spans by conversation ID for one thread, or session ID for the whole session, rather than assume one trace contains the entire job.

Token accounting needs similar care. Usage appears on multiple spans, so adding every token field can count the same usage repeatedly. Anthropic’s instructions say to sum only one source for each request. When a Messages API request has both SDK and API spans, developers should use its API spans and follow the guide’s span-selection rules.

Administrators can manage routing through organization-level endpoints in the beta API reference. Destinations can cover all workspaces, an included list or an excluded list. They can also be enabled or disabled, and expose the latest delivery error and success timestamps. A successful delivery does not erase an earlier error, so the timestamps show whether delivery has recovered.