Telemetry is how an agent shows its working

An agent that can't show its working shouldn't be trusted with more. Telemetry is the record of why it did what it did, and most teams still add it last.

At school the right answer was never enough. You had to show your working. Agents need the same rule. Telemetry is the data a running system sends back about itself: what it did, when, how long it took and what it cost. For an agent the most useful form is the trace, a timeline of one run from the request coming in, through every model call and tool call, to the answer going out. Each step is a span, nested under the step that started it, so you can read a run the way a teacher reads the margin of an exam paper.

Ordinary software, given the same input, does the same thing every time. Anthropic's engineers, writing about their multi-agent research system, note that agents "are non-deterministic between runs, even with identical prompts", and say that adding full production tracing let them work out why agents failed. Without a trace, a bad answer is just a bad answer. With one, you can see the agent picked the wrong tool on step three and never recovered.

Researchers at CSIRO's Data61 go further in their AgentOps taxonomy, arguing you should also record goals, plans, guardrail decisions and memory. Regulators are heading the same way: Article 12 of the EU AI Act requires high-risk systems to technically allow automatic logging of events over their lifetime. For product people, the useful part is what sits on top of the logs. A trace tells you what happened. Evaluations run across thousands of traces tell you whether it was any good.

The richest telemetry is also the most sensitive, because prompts and tool outputs are often your customers' data. The standards work is catching up. OpenTelemetry's GenAI conventions now define shared span types like invoke_agent and execute_tool, so different frameworks report the same shape, and a May 2026 post on the OpenTelemetry blog is clear that by default "no prompt content or tool arguments are captured". OpenAI's Agents SDK takes the opposite default, recording model and tool inputs and outputs unless you switch it off. Anthropic describes watching decision patterns "without monitoring the contents of individual conversations".

I think Anthropic's stance is the right starting point for anyone in a regulated business. Record the structure of every decision by default, and treat content as something you switch on deliberately and keep for a set time. One caveat: the OpenTelemetry agent conventions are still marked as in development, so expect some names to move.

Before an agent gets any more autonomy, ask the team to show you two traces: one run that went well and one that went wrong. If they can't pull them up in a few minutes, the agent isn't ready for more trust, whatever its eval score says. And decide on day one what goes into the trace and what stays out, because retrofitting that later is far harder than it sounds.