AI news · October 16, 2026
Claude adds OpenTelemetry traces for agent observability
Anthropic added OpenTelemetry traces in public beta, including latency, token use and model IDs, so teams can debug Claude agents in the same tools as other services.
- default content cap
- 0.75MB
- trace standard
- OTel
Anthropic announced agent observability for the Claude platform on October 14. The public beta emits OpenTelemetry traces from the Messages API, Managed Agents and SDKs, including model IDs, token counts, latency and request identifiers. Developers can choose metadata-only or full-content modes, send traces to any OTLP-compatible system and use tools such as Grafana, Honeycomb or Sentry. Anthropic documents a default 0.75MB content cap for a single trace.
This addresses a common gap in agent products: an answer can look wrong while the cause sits in a tool call, retry, large prompt or slow service. OTel is more useful than a provider-only dashboard when a workflow crosses models and services. Full content adds inspection value but also increases data handling and storage needs, so the first setup should keep sensitive text out and prove the trace joins work.
What you can do with it
Try the Claude observability setup with metadata-only traces. Run one agent through a tool error and a retry, then inspect latency, tokens, model ID and request IDs in your OTLP tool. Add content only for synthetic data and keep the 0.75MB cap in your test cases.
Our take
This is a high-value developer feature because failed agent calls are otherwise expensive to explain. It does not improve the model by itself, and full traces can create a new data bill. Start with metadata and add content only where it changes a decision.
Links Try itAnnouncement
Source: Anthropic — Made With Models writes the brief; the reporting is theirs.