Traditional observability tells us what happened inside an application: requests, logs, metrics, traces and errors.

But what happens when the application itself starts making decisions?

An AI agent may call several tools, retrieve different pieces of information, change its plan, delegate work to another agent and eventually produce the wrong result — without generating a conventional application error.

This session explores observability for agentic systems, showing how to trace agent decisions, tool calls, retrieval steps, model interactions and agent handoffs. We'll examine how to answer questions such as: Why did the agent make this decision? Where did the context go wrong? Which tool call caused the failure? And how do we measure whether an agent actually performed well?

We'll also discuss practical approaches to evaluating agent quality, debugging non-deterministic behavior and designing observability into agent architectures from day one.