Anthropic launched a tracing beta on October 14, 2026, for teams running Claude-powered customer assistants and internal workflows. The update covers Messages API requests and Managed Agents sessions, letting developers send execution records to monitoring software. Anthropic’s release notes describe the new export capability.
A trace can link model requests and tool activity into a timeline, helping a team investigate where a customer task slowed down. The export uses OpenTelemetry, a shared format for monitoring data; each recorded step is called a span.
Two decisions matter: which parts of the task to follow, and whether the investigation needs conversation text. Collecting more data does not itself fix the application.
Choose the coverage your question needs
What you need to investigate | Starting point | Important boundary |
|---|---|---|
Model runs and tools hosted by Anthropic. | API tracing, configured in Claude Console. | No application-code change is needed for this hosted view. It does not trace your application’s own tools. |
Requests from your application and tools it runs. | Claude software development kit (SDK) tracing in your application. | Covers SDK requests and its tool runner. Other application work needs its own instrumentation. |
A task crossing your application and the hosted service. | Both tracing paths, where API export is supported. | Combine the views; enabling one does not fill the other’s gaps. |
Server-side export is documented for Claude Platform and Claude Platform on AWS, not every cloud offering. SDK tracing also supports Amazon Bedrock, Google Cloud and Microsoft Foundry. Claude Platform on AWS is distinct from Bedrock.
There is an AWS trace-linking exception: server-side spans start separate traces. They do not honor the traceparent header used elsewhere to connect traces and opt requests out of tracing. Do not assume one joined timeline or identical sampling controls across platforms.
For teams considering coordinated document work, our Managed Agents workflow guide addresses when that design is useful. Tracing answers the next question: what happened while it ran?
Decide separately whether to export conversations
Metadata-only is the default. It records timing, status, identifiers and token counts, plus chosen text such as tool names and descriptions; it is not an anonymity guarantee. Full content adds prompts and responses. Each SDK client and export destination has its own setting.
Our recommendation: use metadata first when investigating delays or failed steps. Consider conversation capture when the question needs the exchange itself, such as an incorrect answer. More conversation text cannot compensate for missing measurements of work inside your application.
Review the receiving service’s access and retention settings before exporting content. Anthropic’s retention documentation says the destination holds its copy under its own controls. A zero-data-retention arrangement with Anthropic does not mean the monitoring provider stores nothing. Export eligibility follows the originating feature; Managed Agents is not eligible for zero data retention.
Full-content exports can be shortened; oversized request records lose system instructions and early messages first. Images and documents are represented by type, not reproduced. Check the truncation flag before treating missing instructions as evidence that Claude never received them.
Use traces as diagnostic evidence, not a complete ledger
Anthropic describes API delivery as best effort: records may be lost, while retries can produce duplicates. Deduplicate using the trace ID and span ID. A successful delivery does not establish a complete history.
The span reference also lists gaps, including denied tool uses and multiagent tools in Managed Agents. A missing tool span therefore does not prove an action was never requested. A completed-step status likewise is not a judgment that an answer was correct.
Token charts need care: the same usage can appear at several trace levels. Do not add a Messages request’s total to its child model totals. Follow the documented aggregation rules; otherwise the chart can overstate usage.
For billing reconciliation, use the Usage and Cost Admin API rather than trace totals alone. Its programmatic endpoints are currently unavailable on Claude Platform on AWS, where Anthropic directs users to the Console’s Usage and Cost pages.
Start with one diagnostic question
A useful pilot should establish whether the records answer a real operational question—not merely whether a dashboard receives data. For example: which step delayed a particular customer request?
Ask an administrator to configure a destination for selected workspaces. Confirm that the receiver meets the documented endpoint requirements; choosing all workspaces also includes future ones.
Use the destination’s Test action and confirm the test span arrives. Then run a representative request and check that its expected steps appear. A connection test alone does not verify coverage.
If the missing step belongs to your application, add SDK or application instrumentation. Keep SDK versions pinned during the beta, when options and span schemas can change.
Expand conversation capture only if the unanswered question requires it. Check settings on both export paths, and judge the customer task’s result separately from the trace status.
No separate export tariff was identified in the checked launch, setup or pricing documentation. That does not establish that export or downstream monitoring is free. Confirm monitoring-provider charges before expanding collection; the reviewed sources do not establish a measured saving.
Methodology: AI-assisted reporting and analysis based on Anthropic’s published sources, checked October 14, 2026. No integration or performance test was performed. The pilot sequence is a recommendation, not a tested result.