Claude Platform Now Exports OpenTelemetry Traces for Agents. What You Get and What to Check Before You Rely on It
Anthropic announced public-beta OpenTelemetry traces from Claude Managed Agents, the Messages API and the Claude SDKs. What the announcement says, which tools it names, and a short checklist for tracing an agent before widening its access.
On October 14, Anthropic announced that the Claude Platform can send OpenTelemetry traces from Claude Managed Agents, the Messages API, and the Claude SDKs to any OTel tool. It is in public beta.
If you run agents, this matters for a plain reason: you cannot widen an agent's access safely if you cannot see what it did. Here is what the announcement says, and what I would check before depending on it.
What Anthropic announced
According to the post, you can follow each step an agent takes: its model calls, tool use, and MCP server calls. You add a destination in the Claude Console, and every agent turn, model call, and tool use arrives in your observability tool as a standard OpenTelemetry trace, with latency, tokens, and IDs.
Anthropic says the traces follow the OpenTelemetry GenAI semantic conventions, so data from Claude and other providers can sit in the same dashboards. It also says that if you already send Claude Code or Cowork traces to an observability tool, agents built on the Claude Platform can show up in the same place.
On setup, the post says Managed Agents and the Messages API need a destination in the Claude Console, while the Claude SDKs export traces automatically.
Which tools it names
The post groups tools by purpose:
- Debugging next to your stack: Grafana Cloud, Honeycomb, Sentry, and Splunk.
- Improving agents with production data: Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, and PostHog.
- Usage across an organization: Databricks Unity Gateway, for token usage by team and unusual activity.
It adds that any tool that accepts a standard OTLP endpoint should work. I have not tested any of these integrations.
Customer quotes
The post includes quotes from Foxglove, Ramp, Sterling Labs and another finance customer. They are vendor-selected testimonials. One says that on some agents it cut token usage by 10 to 20 percent and made sessions 30 percent faster, using its own system that scores traces. I cannot verify that, and it is not a result you should expect.
What I would check before relying on it
These are suggestions, not tested results.
1. Confirm it is beta. The post says public beta. Check the docs for limits before building an audit process on it.
2. Decide where traces go. Traces include model calls and tool use. Check what content they carry, and who can read them in your observability tool.
3. Look for one timeline. The useful test is whether a model call and the workflow around it show up in the same trace in your tool.
4. Track tokens per task. Latency, tokens and IDs are in the traces, so cost per finished task becomes measurable.
5. Review before widening access. The post's own advice is to review traces, then widen access as the agent proves reliable. Write down what "reliable" means first.
6. Keep retention in mind. If traces are evidence for an audit, set retention on your side.
What I would not conclude
Tracing shows what an agent did. It does not stop an agent from doing something. Do not treat observability as a replacement for permission limits.
And do not assume every field you need is in the traces. Check a real trace first.
Frequently asked questions
What did Anthropic announce?
OpenTelemetry trace export from Claude Managed Agents, the Messages API and the Claude SDKs on the Claude Platform, in public beta.
What is in a trace?
Per the post: agent turns, model calls, tool use and MCP server calls, with latency, tokens and IDs.
Which tools work with it?
The post names Grafana Cloud, Honeycomb, Sentry, Splunk, Arize AI, Braintrust, Datadog, Langfuse, LangSmith, New Relic, PostHog and Databricks, plus any OTLP endpoint.
How do I turn it on?
Add a destination in the Claude Console for Managed Agents and the Messages API. The SDKs export automatically, per the post.
Does it follow a standard?
Anthropic says it follows the OpenTelemetry GenAI semantic conventions.
Are the customer results verified?
No. They are quotes selected by the vendor.
Does tracing make agents safe?
No. It gives visibility. Access limits are separate.
Sources
- Anthropic, Agent observability on the Claude Platform: all feature, setup and tool claims (October 14, 2026).
- TLDR AI, October 15 issue: discovery pointer only.
Related reading
- Anthropic Published Its Claude Unintended-Actions Report. What It Means If You Run Agents With Real Access
Anthropic's October 9 report describes Claude exploiting a server flaw, submitting a real form, using access tokens to skip a fee, and using URL shorteners to beat a fetch limit. Here is what it says, and the boundaries builders should set before an agent touches live systems.
- Google OAuth Now Supports DPoP for Refresh Tokens. What It Binds, What It Leaves Alone, and the Agent Delegation Gap
Google added DPoP (RFC 9449) support to its OAuth platform, but only for refresh tokens. What the announcement and adoption guide say, what stays a bearer token, and what Google says about multi-agent and MCP chains.
- Google's AQuA Reads Your Production Agent's Traces. What It Does, What It Costs, and What Google Has Not Shown
Google open-sourced AQuA, an ambient quality agent that samples production sessions, clusters failures and diagnoses them against deployed source. What Google's post says, with its own numbers and limits.
The build log
Follow the work.
New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.
Follow via RSS