Tracing the Mind of the Machine: Observability for AI Agents

AI agents have evolved beyond LLM chatbots; they possess the ability to plan, reason, and act autonomously. However, as their autonomy increases, understanding how they make decisions becomes more challenging. Traditional methods of observability—such as metrics, logs, and traces—capture outcomes but do not reveal the underlying reasoning. This session will explore how AI Agent Observability can shed light on the decision-making process by collecting and analyzing agent traces. We will discuss emerging standards like the Model Context Protocol (MCP), which provides structured and shareable traces of agent interactions with tools, memory, and goals. This gives developers unprecedented visibility into the "why" behind an agent’s actions. Attendees will learn how to: capture reasoning traces to reconstruct an agent’s decision pathway, use MCP as a backbone for standardized observability across different agents and frameworks, detect deviations, hallucinations, or goal misalignments before they impact users, and build trust and reliability in production environments by making black-box reasoning auditable and explainable.