Is your AI app working well until it starts failing on complex tasks? It’s a common pain point that happens because LLMs tend to forget instructions, get confused, and, over time, their performance degrades.

The problem is not the models, but how you feed them information. This is the core challenge of Context Engineering: actively managing an LLM’s short-term memory to ensure it has the right information available at the right time. A bigger context window is not a magic solution. Mastering its management is the key to building robust and efficient AI systems.

In this session, we’ll cover the 3 silent killers of LLM performance: context rot, context pollution, and context confusion. We’ll also look at a reliable mental model for managing context in multi-agent systems, and how to design a hierarchical action space to prevent tool-related confusion.

This session is for software developers and AI engineers building LLM-powered apps who are struggling with agent reliability and performance on complex, long-running workflows.

Jfokus

A conference talk transcript covering context engineering for LLMs. The speaker explains why the context window is the LLM's working memory and why stuffing it with irrelevant tokens degrades quality, increases cost, and raises latency due to quadratic attention complexity. Three core problems are identified: context rot (stale information), context confusion (conflicting instructions from system vs. user prompts), and context pollution (irrelevant data drowning out key signals). Additional challenges include tool overflow in multi-agent systems and the 'lost in the middle' effect. Mitigation strategies discussed include compaction, summarization, agent specialization, modular skills, and context routers with intent classifiers and context assemblers.

Stop Stuffing Your LLM's Context Window, by César Soto Valero