Claude Platform adds tracing for API requests and agents

Anthropic launched OpenTelemetry export for Claude Platform in beta on October 14, 2026. The release note covers Messages API requests and Claude Managed Agents sessions, with SDK spans available in Python, TypeScript, C#, Go, Java, PHP and Ruby. Developers can send that activity to an observability service they already use.

The practical question is where a slow or failed task spent its time. A customer support application might wait on a database lookup, a model response or a tool running elsewhere. Collecting more records helps only when the team can connect those steps and understands what the records leave out.

Choose the layer you need to inspect

Anthropic’s coverage overview distinguishes the two tracing layers. API spans expose internal model requests, agent turns and hosted tools. SDK spans cover client tools and self hosted sandbox activity. A span records one step.

The SDK guide requires a tracer provider registered before the first call being traced. The application’s exporter sends the spans. A streaming request’s span finishes when the application consumes or closes the stream, so code that stops reading early must close it. Another instrumentation library can produce duplicate request records.

For API export, the setup guide requires an organization admin to configure a destination and its workspace scope. The receiver must accept protobuf OTLP over HTTPS on port 443. A collector listening only on its usual port 4318 needs an appropriate HTTPS endpoint. Developers can inspect destinations, with narrower workspace administration controls. The Admin API and CLI provide alternative configuration routes with appropriate administrative access.

API export supports the Claude API and Claude Platform on AWS.

SDK tracing also supports Amazon Bedrock, Google Cloud and Microsoft Foundry.

That split should shape the rollout. Identify whether the missing evidence sits in your application or inside the provider before changing both systems. Assign responsibility for the application exporter and the destination, then agree on a sample request the owners can locate.

Decide what may leave the application

The default Metadata only mode includes timing, status, identifiers, token counts and chosen agent or tool names and descriptions. Full content adds prompts and responses. The content settings normally cap content at 0.75 MB per span and truncate larger records. Images and documents appear by type, rather than as complete copies.

A useful starting point is metadata in a limited workspace. For the support application, that can help investigate which step took time without immediately copying customer conversations into another system. Expand content capture only for an explicit diagnostic question, with a defined audience and retention period.

Metadata still needs review. The SDK exports conversation identifiers in both modes, and Anthropic warns against putting secrets or personal data in them. The application’s propagator can also send baggage with trace context. Review those settings rather than assuming the content switch controls every exported field.

The destination stores received data under its own access and retention rules. Zero data retention eligibility depends on the underlying Claude feature, and Managed Agents is excluded. Adding a telemetry destination does not extend Anthropic’s retention terms to that external store.

Keep token totals separate from the trace tree

Anthropic’s counting guidance says to count one source and one span level per request. When both SDK and API spans exist, use the API spans. For Messages API totals, choose request spans or their model request children, then account for compaction separately. For Managed Agents, sum model request and outcome evaluation spans.

The span reference says input usage already includes cache reads and writes. Adding those cache fields again inflates the total. Compaction usage is outside the ordinary request input and output totals. With SDK records, add the compaction usage attributes. With API records, use compaction spans without adding the same attributes again.

There is a separate duplication problem in delivery. API spans may arrive more than once, so deduplicate using trace and span IDs. Delivery is best effort, lost spans cannot be resent, and a Healthy destination means its last delivery succeeded.

These details change what a cost dashboard can claim. Keep a documented counting rule beside the query, and compare its output with the provider’s usage records before using it for budgets. A missing record should prompt investigation. It cannot establish that an operation never happened.

Keep Claude Code monitoring in its own setup

Claude Code has separate OpenTelemetry monitoring for metrics, events and optional traces. Its tracing controls are configured for Claude Code and are off by default. Teams already collecting those signals should review the Platform integration separately instead of treating the October 14 announcement as a new switch for their coding terminal.

For applications using Managed Agents, ByteForward’s guide to dynamic workflows covers the related questions of budgets, permissions and checking that every input was handled. A trace can help investigate a run, while the application still needs acceptance criteria for the result.

Validate one known request before relying on the dashboard

Start with a small request using sample data you can safely inspect. Write down the expected tool call and result before running it. Then check three things in the receiving service.

• Can the reviewer follow the application request through the expected model and tool steps without mistaking a duplicate for another action?
• Do the displayed fields match the intended content policy, including names, identifiers and any text that should have remained private?
• Does the usage calculation count that work once, with any compaction accounted for and a clear way to notice gaps?

Repeat the check with a deliberately interrupted stream and a controlled tool failure. Those are proposed evaluation steps, not tests ByteForward has performed. This article is based on the release note and documentation. The beta’s value will depend on whether it makes a real failure easier to explain without creating a second uncontrolled store of application data.

Illustrative photograph by Lightsaber Collection on Unsplash under the Unsplash License. Published in 2023.