OpenTelemetry just standardized LLM tracing. Here's what it actually looks like in code.

OpenTelemetry has published experimental GenAI Semantic Conventions that standardize how LLM calls, agent invocations, and tool executions are traced, replacing the fragmented vendor-specific formats. The spec defines three span types (chat, invoke_agent, execute_tool), standardized attribute paths like gen_ai.agent.name and gen_ai.tool.name, and guidelines for content recording. The author shares practical implementation lessons, including common mistakes like wrong attribute paths and span naming formats that make traces invisible to GenAI-aware backends.