Claude OpenTelemetry Tracing for Business Agents

Anthropic put OpenTelemetry tracing for the Claude Platform into public beta on October 14, 2026. Agents built on Claude Managed Agents, the Messages API or the Claude SDKs can now send a record of every model call, tool use and agent turn to the monitoring tool a company already uses. That covers Grafana Cloud, Sentry, Datadog, Langfuse, LangSmith and any other tool that accepts the standard OpenTelemetry (OTLP) format.

For a business, this answers a plain question that used to need custom code: what did the agent actually do, step by step, and what did each step cost?

We read Anthropic's announcement, the OpenTelemetry overview and the API tracing setup page on October 17, 2026. Everything about Claude below comes from those pages. We have not run a production workload through the beta for this post.

If you run an agent on the Claude API that touches customer data, money or outside systems, turn this on and start with the default setting, which sends timing, token counts and IDs but not the prompts and replies. The feature is for agents built on the Claude API and SDKs. Claude Code, the Agent SDK and Cowork have their own monitoring pages.

A trace is the full record of one request as it moves through your systems. A span is one step inside it, such as one model call or one tool use. Anthropic's spans follow the OpenTelemetry conventions for generative AI, so tools that already understand those conventions can read them without extra setup.

In practice you open your monitoring tool, find the request a customer complained about, and see the chain: your app received a message, Claude was called, it ran a search tool, it called one of your own tools, it was called again, and it answered. Each step shows how long it took and how many tokens it used.

Anthropic offers two routes, and you can use both.

1. 1API tracing. An organization admin adds a destination in the Claude Console: the address of your monitoring tool and the credentials it needs. The endpoint has to accept OTLP over HTTPS on port 443. After saving, a Test action sends a test span so you can confirm delivery. API tracing covers each model run inside a request, Managed Agents turns, and tools the API runs for you, such as web search.
2. 2SDK tracing. A developer sets it up in code with the Claude SDK. It needs no admin role. It covers tools your own code runs through the SDK's tool runner, tools in a self-hosted sandbox, and other API calls such as file uploads.

With both on, the SDK span and the API span for a request show up as one trace. A request can also nest inside your own application's trace, because the API reads the standard W3C traceparent header. That is how you link an agent step to the customer, order or ticket that triggered it.

If you connect your own MCP server through the Messages API's MCP connector, the API passes the trace on. Every request it sends to your server carries a traceparent in the message metadata, and a tool call points at the matching tool-use span. If your MCP server records its own spans, they nest under Claude's. For anyone running custom MCP servers, that removes a lot of guesswork about which agent request caused which database query.

Each destination has a content mode. Metadata only is the default: IDs, timing, status, token counts, and names you chose, such as tool and MCP server names. Full content adds the prompts and responses.

Full content is what lets you read a failed conversation or build test cases from real traffic. It also means customer messages now live in your monitoring tool, under that tool's access and retention rules. Anthropic says so directly. A few details matter here:

• If your organization has a HIPAA, FERPA or K-12 data restriction with Anthropic, destinations export metadata only no matter what you pick. SDK spans are not affected by that rule.
• With zero data retention, Full content still works for Messages API requests. Managed Agents is not eligible for zero data retention.
• Each span holds up to 0.75 MB of content by default. Longer content is cut, and the span is marked as truncated.

Anthropic describes delivery as at least once and best effort. A span can arrive twice, and a span can be lost and is not sent again. A destination is switched off if it redirects, points to a non-public address, or fails for 24 hours straight. Anthropic also says exported spans are not a complete record of API usage and points to its Usage API for tracking each key.

So use traces to debug and to measure. Do not use them as the only record of what an agent did to a customer account. Keep that record in your own system.

These are designs we would build. None of them describes a client result.

1. 1Trace review before widening access. Anthropic's own framing is to review traces and give an agent more access as it proves reliable. We would make that a routine: every week, pull the tool calls for one agent, look for wrong tools, retries and failed steps, and only then add the next permission. It pairs with our guide on AI agent permissions and human approval.
2. 2Cost per workflow. Tag each agent run with the workflow it belongs to, such as invoice intake or lead follow-up, and total token use per workflow from the spans. Anthropic gives specific rules for counting each token once, because usage repeats on several span levels. Follow them, or the numbers will be inflated.
3. 3One timeline across your own tools. For agents that call your MCP servers, connect the traceparent so a slow answer can be traced to the slow database query behind it.
4. 4Test cases from real runs. With Full content turned on for a test workspace only, save the runs that went wrong and replay them before each prompt or tool change. Our post on testing AI agents before production covers what to check.

1. 1Pick one monitoring tool. If you already use Grafana, Datadog or Sentry for the rest of your systems, send agent traces there so they sit beside your logs.
2. 2Start with Metadata only on all workspaces.
3. 3Turn on Full content only where you have decided who can read customer conversations and how long they are kept.
4. 4Remove duplicate spans by trace ID and span ID if your tool does not do it.
5. 5Alert on a failing destination. A silent monitoring gap is worse than none, because people assume it works.
6. 6Keep your own action log for anything that changes money, records or customer accounts.

If your team only uses ready-made Claude apps and has built nothing on the API, this beta has nothing for you to set up. If you run one low-stakes agent a few times a day and read its output anyway, a monitoring tool adds cost and setup for little gain. If you do not have any monitoring tool yet, adding one only for this is a project of its own, and it may be better to start with simple logging. And because it is a beta, plan for the setup details to change.

OrchestriAI does not resell Anthropic or any monitoring product. We build the agents and the plumbing around them. AI agent systems cover what an agent may do and how you check it. MCP development covers servers that record their own spans and accept the traceparent. Systems integration connects traces to the orders, tickets and records they belong to. For the hardening side of a custom server, see securing a production MCP server. If you are not sure whether an agent is the right tool for your process at all, start with the free automation audit.

OrchestriAI is an independent implementation provider. Product names are used for identification only. OrchestriAI is not affiliated with, endorsed by, or sponsored by Anthropic or any observability vendor named here. This article summarizes Anthropic's public documentation as of October 17, 2026. The feature is in beta, so confirm current behavior in Anthropic's docs before relying on it. This is not legal, privacy or compliance advice.