Agent Observability on the Claude Platform

Developers can send OpenTelemetry traces from Claude Managed Agents, the Messages API, and Claude SDKs to any OpenTelemetry tool and follow each step an agent takes. claude.com's report says a destination in the Claude Console sends every agent turn, model call, and tool use as a standard OpenTelemetry trace with latency, tokens, and IDs. Claude Platform agents can show up in the same tool that already stores Claude Code or Cowork traces. The export is in public beta.

This piece explains the trace fields, the Console and SDK export paths, and the observability tools that receive them. It is for engineers who run agents on the Claude Platform, and for platform and security teams that need token usage by team and a way to flag unusual activity.

Agent Observability on the Claude Platform: what actually changed

The Claude Platform now exports OpenTelemetry traces from Claude Managed Agents, the Messages API, and Claude SDKs. Each standard trace carries the agent turn, the model call, and the tool use, plus latency, tokens, and IDs. Teams trace model calls, tool use, and MCP server calls, review them, and widen access once the agent proves reliable. The export is in public beta. Managed Agents and the Messages API require a destination in the Claude Console. Claude SDKs export traces automatically.

Teams that already send Claude Code or Cowork traces can keep one destination. Agents built on the Claude Platform then show up in that same observability tool. Claude traces follow the OpenTelemetry GenAI semantic conventions, so Claude data and data from other providers work in the same tools and the same dashboards. The transport is a standard OTLP endpoint. Any OpenTelemetry tool that accepts OTLP can take the traces, which covers products beyond the named debugging, evals, and token-usage tools.

Agent Observability on the Claude Platform: how it works

For Claude Managed Agents and the Messages API, setup is a destination in the Claude Console. After it is added, every agent turn, every model call, and every tool use arrives in the observability tool as a standard OpenTelemetry trace with latency, tokens, and IDs. Claude SDKs export traces automatically. Both paths are in public beta. The same trace is the access record: model calls, tool use, and MCP server calls stay visible, and wider access follows as the agent proves reliable.

Grafana Cloud, Honeycomb, Sentry, and Splunk place agent traces beside logs and metrics, which shows the service a slow or failed step came from. Arize AI, Braintrust, Datadog, Langfuse, LangSmith by LangChain, New Relic, and PostHog turn real traces into evals, catch quality issues, and measure whether a change helped. Databricks Unity Gateway lets platform and security teams see token usage by team and flag unusual activity. Any other OpenTelemetry tool that accepts a standard OTLP endpoint can take Claude traces.

Agent Observability on the Claude Platform: why it matters now

Trusting an agent with more access starts with tracing its model calls, tool use, and MCP server calls, then widening access as the agent proves reliable. Debugging uses that record beside the rest of the stack. In Grafana Cloud, Honeycomb, Sentry, and Splunk, OpenTelemetry traces also sit beside logs and metrics, which shows which service a slow or failed step came from during each agent turn. Those dashboards follow the OpenTelemetry GenAI semantic conventions, so other providers can share them.

Ramp reported these production results. David Traina, Senior Software Engineer, said Claude powers some of Ramp's most capable internal and customer-facing agents. Ramp built Reflect, an agentic system that scores Langfuse traces, finds recurring failures, and proposes fixes. On some agents, Ramp has cut token usage by 10-20% and made sessions 30% faster. Traina said, "Reflect can only improve what it observes." Native OpenTelemetry export from the Claude API can feed high-fidelity instrumentation directly into that improvement loop, with no extra steps.

Agent Observability on the Claude Platform: who is affected

Roman Shtylman, Co-Founder and CTO of Foxglove, said Foxglove builds agents that help robotics teams debug their robots. OpenTelemetry export lets Foxglove analyze agent interactions in PostHog alongside product usage data, evaluate how the agents perform in practice, identify where they fall short, and prioritize improvements for robotics teams. Hyon Lee, CTO of Sterling Labs, said Sterling Labs tunes its systems like a race team tunes its engine. Agents analyzing live markets 24/7 use Langfuse as the pit wall. Every trace streams straight in with no custom instrumentation, so his team can focus on the edge, not the wiring.

Ankush Sharma, Head of Infrastructure, said agents that touch a company's books have to show their work, the same way an accountant does. Those agents run on the Claude Platform with telemetry in Grafana Cloud. Native OpenTelemetry export puts each model call in the same trace as the reconciliation or journal-entry workflow it ran inside, with the prompt version, tool calls, latency, and token spend. When an auditor asks what an agent did and why, they answer from one timeline.

Agent Observability on the Claude Platform: what to watch

OpenTelemetry traces on the Claude Platform are in public beta. Managed Agents and the Messages API emit traces only after a destination is added in the Claude Console. Claude SDKs export traces automatically. Teams sending Managed Agents or Messages API traces to Grafana Cloud, Honeycomb, Sentry, Splunk, or any standard OTLP endpoint add that destination first. Each trace follows the OpenTelemetry GenAI semantic conventions, carries latency, tokens, and IDs, and can share a tool with Claude Code or Cowork traces.

On some Ramp agents, Reflect has cut token usage by 10-20% and made sessions 30% faster. David Traina said Reflect scores Langfuse traces, finds recurring failures, and proposes fixes, and Reflect can only improve what it observes. Native OpenTelemetry export from the Claude API can feed high-fidelity instrumentation directly into that improvement loop, with no extra steps. Foxglove analyzes agent interactions in PostHog beside product usage data. Sterling Labs streams every trace into Langfuse with no custom instrumentation for live markets 24/7.