If you have used an LLM agent, you probably have noticed that at the start of a new conversation the agent is smart, helpful and to the point. But as the conversation goes on it starts to make more and more silly mistakes, forgets important facts, or even gets stuck entirely. This 'context rot' is inherent to how LLMs work, but production AI agents must be able to remember key information and not go off the rails when working on a bigger task. So how do we make our agents more reliable?

This session introduces Context Engineering as an evolution of prompt engineering: the discipline of dynamically constructing the right context at the right time across an agent’s lifecycle.

In this session, you will learn how to practice Context Engineering by writing, selecting, compressing, and isolating information to keep agents accurate and stable over time. The session will also cover how to design agent memory, combining short-term state with long-term memory using patterns like rolling summaries and structured user profiles.