A pod can look perfectly healthy on every dashboard while the agent inside it loops through [4,000] tool calls overnight, hands a task to a second agent that silently drops the context, and burns [$X] in tokens before anyone notices. These are behavioral failures, not infrastructure failures, and the request/response trace model most teams run today was never designed to capture them.

This session shows how to instrument a multi-agent system on Kubernetes with OpenTelemetry so that decisions, not just requests, become observable. Using an open-source agent framework, the OpenTelemetry Collector, and an open-source backend, we walk through the GenAI semantic conventions as they stand today, then go past them: propagating context across MCP tool calls and agent-to-agent handoffs, modeling loops and retries as span trees, attributing tokens and cost to the agent step and Kubernetes namespace that caused them, and correlating agent spans with inference-gateway and GPU signals to separate “the model is slow” from “the agent is thrashing.”

Attendees will leave with a reference instrumentation pattern, Collector pipeline configuration for aggregating and sampling agent telemetry, and a clear picture of where the conventions still fall short and how to feed those gaps back to the project.