AILet’s talk
Back to AI updates
AnthropicNew features

Agent observability on the Claude Platform

This is a developer/platform-team tool: it lets teams building agents on Claude see exactly what each agent call did (latency, tokens, tool use) inside whatever monitoring tool they already use, which matters for tracking cost and reliability, not for a non-technical end user.

Read the original source 2 min read
Diagram. On the left, a panel titled 'OpenTelemetry traces' shows a timeline of nested spans: Agent, API call, Model, Tool use, Model, Model and Database. A legend colours the bars by source: Your app (grey), Claude (orange) and Other providers (blue). Lines connect this panel to three cards on the right. 'Debug failures: Filter by tool and error status to find the root cause' shows the attribute gen_ai.tool.name, status and a red bar labelled Error. 'Track costs: See token usage per request and per session' shows gen_ai.usage.input_tokens with bars labelled 48.2k, 30.9k and 18.4k. 'Improve agents: Flag real traces in your backend and turn them into evals' shows trace bars with an arrow pointing to 'Evals'. Above the cards is a row of observability tool logos, including Splunk, followed by a plus tile. A dashed arrow loops from the cards back to the traces panel.
Anthropic's diagram of OpenTelemetry tracing on the Claude Platform. Agent, model-call and tool-use spans from Claude appear next to spans from your app and other providers. You can then use them in observability tools to debug failures, track token usage and turn real traces into evals. The feature is in public beta. Managed Agents and the Messages API need a trace destination set up in the Claude Console, while the Claude SDKs export traces automatically. Values in the diagram are examples. ↗

What changed

Source summary

Anthropic added OpenTelemetry (OTel) tracing for Claude Managed Agents, the Messages API, and the Claude SDKs: each agent turn, model call, and tool use is exported as a trace with latency, token counts, and IDs, following OpenTelemetry's GenAI semantic conventions so it integrates with existing observability tools. It is in public beta on the Claude Platform; Managed Agents and the Messages API require adding a trace destination in the Claude Console, while the Claude SDKs export traces automatically.

Why it matters to your work

Practical interpretation

This is a developer/platform-team tool: it lets teams building agents on Claude see exactly what each agent call did (latency, tokens, tool use) inside whatever monitoring tool they already use, which matters for tracking cost and reliability, not for a non-technical end user.

When this could help

Relevant when you're running agents in production on the Claude Platform and need to debug performance or track token usage by team across an organisation.

Anthropic added OpenTelemetry (OTel) tracing so every agent turn, model call, and tool use exports a trace with latency, tokens, and IDs, following OTel's GenAI semantic conventions. Traces work with named partner tools (including Grafana Cloud, Honeycomb, Sentry, Splunk, Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog, and Databricks Unity Gateway) and with any other OpenTelemetry tool that accepts a standard OTLP endpoint.

What to check

  • Public beta; Managed Agents and Messages API tracing requires manual configuration of a trace destination in the Claude Console. Source ↗

Sources & provider guides

Official documentation checked 23 Oct 2026. SCALA prompts and workflows are suggestions, not claims of hands-on testing.

Source & availability

This is a summary of a published provider announcement, with practical context from SCALA’s feed.

Plan, region, workspace and rollout conditions may affect access in Australia. Check the original source for current availability.

claude.com

This update is primarily for technical teams.