Anthropic has rolled out OpenTelemetry tracing across its Claude Platform, giving developers a standardized way to see exactly what their AI agents are doing at every step. The feature entered public beta on October 14, 2026, and covers Claude Managed Agents, the Messages API, and the Claude SDK.
What the new tracing actually does
The spans Anthropic generates comply with OpenTelemetry GenAI semantic conventions, which is the industry-standard schema for describing AI system behavior. That means each trace automatically includes latency figures, token counts, model names, tool usage records, and automated status reports.
Developers who want deeper insight can turn on an optional full-content mode. When enabled, that mode captures the actual text of prompts, responses, and tool inputs and outputs, which makes reproducing and diagnosing failures considerably faster.
Routing the traces requires minimal configuration. Developers can point an OTLP destination inside the Claude Console, or register a tracer provider directly inside any of the supported SDKs. Anthropic lists Python, TypeScript, Go, Java, C#, PHP, and Ruby as compatible languages.
The traces are designed to work alongside logs and metrics inside tools developers already use. Grafana Cloud, Sentry, Splunk, and New Relic are all compatible destinations, so teams do not need to adopt a new monitoring stack to take advantage of the feature.
Why end-to-end visibility matters here
The coverage is broader than it might initially sound. A single trace can span application code, Claude model calls, platform-run tools, and interactions with downstream Model Context Protocol servers. MCP is Anthropic’s protocol for connecting AI agents to external data sources and services, so tracing through it closes a gap that previously made multi-step agent pipelines opaque.
Anthropic says this build extends observability work it had already done for Claude Code and Cowork, meaning agents built on those surfaces should fit into the same dashboards without additional instrumentation.
Context within a busy product cycle
The OpenTelemetry announcement lands less than two weeks after a pair of other platform updates. On October 9, 2026, Anthropic introduced dynamic workflows for Managed Agents, and on October 7, it launched Claude Haiku 5.5.
By conforming to OpenTelemetry’s GenAI semantic conventions rather than building a proprietary telemetry format, Anthropic is also making a deliberate interoperability bet. Developers who already use OpenTelemetry for the rest of their stack can treat Claude as just another instrumented service, which lowers the integration cost significantly.
The public beta label means the feature is available but may still change. Teams building production systems around the traces should account for the possibility that span schemas or configuration interfaces could shift before a stable release.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.