OpenTelemetry for AI Agents: What to Track as Names Change

Why a latency panel can go flat

IN SHORT A metric name your dashboards rely on can change under you, because the spec behind it is still moving.

IN PLAIN WORDS

Telemetry is what a running system reports about itself: timings, counts, logs. OpenTelemetry is an open standard for it. Each spanSpan One timed step of work, such as a model call or a tool call, with a name and labelled values called attributes. See the glossary in a traceTrace All the spans of one run, linked parent to child, so you can see which step started which. See the glossary carries attributes, like the model name.

That change landed on 6 October 2026. KubeCon North America, the big cloud conference, runs 9 to 12 November in Salt Lake City: a good moment to decide what to instrument next.

I built and ran 46 internal-ops agents at Integra. I also built a separate 11-agent platform for a private-equity fund. This isn’t a report from either one: it’s the mapping I’d apply today to the gen_ai.* names.

ADAM

ADAM says: Think of a shared paper form: every office fills in the same printed boxes. Where it breaks: this form has no edition number, so you write the commit down yourself.

Think of a shared paper form: every office fills in the same printed boxes. Where it breaks: this form has no edition number, so you write the commit down yourself.

▼

?Coming up next: So where do these shared names come from?

IN SHORT They’re OpenTelemetry’s shared names for AI telemetry. Since June 2026 they live in their own repository, apart from the core conventions.

These semantic conventionsSemantic conventions OpenTelemetry’s agreed names for spans, attributes and metrics, so every tool means the same thing by the same name. See the glossary give each model call a known span name and known attributes, under gen_ai.*. Any compatible tool can then count tokens per model.

They moved this year. Core v1.41.0 (28 April 2026) and its v1.41.1 patch (11 May) were the last core releases with the GenAI pages. Since v1.42.0 (12 June 2026), a “Moved” notice points to semantic-conventions-genai.

None of the GenAI conventions there is stable. Every GenAI span, metric and attribute I looked at is in Development statusDevelopment status OpenTelemetry’s earliest maturity level. Names can still change or break, and the project advises against relying on them in production. See the glossary. There’s no tag, no release, and the README’s Schema URL section just says “TODO”.

The manifest, the file that describes the repository, gives a placeholder schema URL, gen-ai-dev/1.42.0-dev. On 7 October it returned a 404. Only the stable attributes borrowed from core, like error.type and server.address, are guaranteed stable.

IN SHORT For model calls, gen_ai.client.inference.duration, merged on 6 October 2026. It’s the latest of several breaking changes since August.

It came with PR #521 (a pull request: a proposed change). The old metric stays for other operations. But the metrics page now says it “SHOULD NOT be reported for inference operations”.

Each change can quietly empty a panel built on the old name, once that name stops arriving. That’s not a complaint: Development status exists so names can improve before they freeze.

ITEM GET

Name the commit

With no release, following the GenAI conventions means following one commit of them, and saying which.

Sources, read on 7 October 2026: semantic-conventions-genai, commit for #521, metrics, manifest, core v1.42.0 notice, maturity levels.