On 14 October 2026 Anthropic released OpenTelemetry tracing in public beta across the Claude platform. It covers Claude Managed Agents, the Messages API and the Claude SDK. The search term to watch is “AI agent observability”, which is moving from engineering jargon to a line on buyer checklists.
What it actually does
Every step an AI agent takes can now be recorded as a trace: how long each call took, how many tokens it used, which model ran, which tools were called and whether they succeeded. A single trace can follow a request from your own application code, through the model, into the tools the agent used and out to any connected services. An optional full-content mode also records the prompts and responses themselves.
Traces use the standard OpenTelemetry format, so they can be sent to monitoring tools teams already use, such as Grafana Cloud, Sentry, Splunk or New Relic. No new monitoring stack is needed. It is a public beta, so the data format and configuration may still change.
Why this matters if you are not an engineer
AI features fail differently from normal software. A normal bug throws an error. An AI agent can quietly give a wrong answer, loop on a task, or run up a bill. Without tracing, the first you hear about it is a customer complaint or an invoice. With tracing, someone can see which step went wrong and reproduce it.
The same applies to cost. Token counts per step are exactly the data you need to answer “why did our AI bill double?” without guessing.
One caution
Full-content tracing stores prompts and responses. If those contain customer, patient or financial data, that is personal data you now hold somewhere else. Decide what gets logged, where it is stored and for how long before switching it on, particularly in regulated work.
So what
If you are commissioning an AI product, add three questions to your brief: how will we see what the agent did, who watches the cost per task, and what is logged about our users. A supplier who cannot answer clearly is building a demo, not a product.
Observability is part of how we build AI products and run AI-assisted development at BuildApps. If you are scoping an AI feature and want monitoring designed in from day one, start a project.