OpenTelemetry push for GenAI spans
Carter Theogene laid out GenAI semantic conventions for OpenTelemetry — a gen_ai.* span schema and a four-step production checklist to standardize LLM/agent traces posted. The goal: predictable dashboards and cross-provider observability for model spans, prompts and tool calls.
OpenTelemetry’s GenAI span spec prescribes span names like {gen_ai.operation.name} {gen_ai.request.model} and recommends SpanKind=CLIENT for model calls (INTERNAL only for in-process models). opentelemetry.io The conventions add explicit attributes — gen_ai.request.model, gen_ai.response.model and gen_ai.provider.name as a provider discriminator — and introduce client/server metric names such as gen_ai.client.token.usage and gen_ai.server.time_to_first_token for cost and performance attribution. opentelemetry.io The repo is marked “Development” and includes a transition policy: instrumentations using pre-v1.36.0 MUST keep defaults while maintainers added an OTEL_SEMCONV_STABILITY_OPT_IN env var to opt into gen_ai_latest_experimental behavior. opentelemetry.io A focused GitHub proposal (issue #2664) expands agentic semantics with attributes for tasks, actions, agents, teams, artifacts and memory links to represent multi-step agent workflows in traces. github.com Ecosystem packaging and instrumentations are already appearing: opentelemetry-semantic-conventions-ai 0.4.15 was published to PyPI on Mar 2, 2026, and opentelemetry-instrumentation-openai-v2 provides auto-instrumentation to capture prompts, completions and related events. pypi.org Commercial vendors and cloud SDKs are integrating the spec — OpenLLMetry conventions merged into OpenTelemetry (noted by Dynatrace on Jan 21, 2025), Elastic published model-monitoring guides (June 21, 2023), and Vercel’s AI SDK documents OpenTelemetry telemetry integration. community.dynatrace.com Hands-on guides and tutorials illustrate operational patterns that the conventions enable: trace-correlated tool-call spans for replayable debugging, token-cost metrics for bill-of-materials cost attribution, and single-trace failure analysis for multi-agent loops in practical posts from OneUptime and nnode.ai in early 2026. oneuptime.com
