Dove sono finiti i tuoi token? Tracciare un sistema multi-agente con OpenTelemetry

A multi-agent workflow that calls tools and other agents is, for all practical purposes, a distributed system: latency that piles up, token costs that explode, failures that cascade. And yet end-to-end visibility is often missing. In this talk I show how to instrument an agentic system with OpenTelemetry and the GenAI semantic conventions: the official span tree (invoke_agent, then a chat span for each LLM call, then execute_tool for each tool), the gen_ai.* attributes for tokens, model and finish reason, and end-to-end latency. The question to answer live: why did this workflow give a wrong answer after spending more time and more money than expected? Cost is not a standard attribute: you derive it from tokens and pricing, and I show how. The same span tree becomes an audit log: who called which tool, with which tokens, when and why it failed. Three honest notes: the GenAI semantic conventions are still experimental, and I present them as a standard that is still settling. I am building the demo for this talk, bringing observability discipline into a domain that is new to me. And if you capture prompts and responses, that text ends up in the spans sent to your backend: you need redaction or filtering before exporting it.