Traditional cloud-native monitoring wasn't built for autonomous AI agents. A pod might show perfectly healthy CPU and memory while the agent inside quietly burns $5,000 over a weekend looping through failed API calls.

As systems evolve from simple prompts into complex Agent-to-Agent (A2A) and Model Context Protocol (MCP) workflows, standard metrics leave platform teams debugging in the dark. Real-world failures demand a shift toward observing AI decision paths rather than just system health.

This technical deep-dive explores how to implement OpenTelemetry GenAI semantic conventions to solve this blind spot. Attendees will learn to instrument agents across MCP boundaries, manage context propagation for multi-agent handoffs, and achieve step-level token and cost attribution per tenant. Ultimately, we will demonstrate how treating evaluation signals as telemetry prevents runaway costs and builds trust in scalable autonomous systems.