The State of Context Engineering

At AGNTCon, a talk titled "Context Fails & Fixes" was given, serving as an overview of the current state of context engineering. As model creators pursue new capabilities, they inadvertently introduce challenges for users and harness builders. This talk covers the emergence, evolution, and future prospects of context engineering.

The speaker mentioned their upcoming book, "The Context Engineering Handbook," which will address the ongoing challenges in AI instruction and communication. Despite the rapid aging of AI textbooks, the speaker expressed optimism about addressing these issues.

Context engineering remains difficult due to the unique attributes and construction of models, which differ from people. While assumptions can be made about their thinking, understanding the models' inner workings is crucial for dealing with their failure modes.

The talk highlighted four key failure modes: context poisoning, distraction, confusion, and clash. Context poisoning occurs when irrelevant or incorrect information is introduced into the context, causing the model to become fixated on it. Context distraction happens when the model focuses too much on the context, overshadowing its base knowledge.

Context confusion arises when extraneous facts are added to the context, influencing the model's behavior in unintended ways. Finally, context clash occurs when different tools, skills, or information within the system conflict with each other, leading to unexpected results.

The speaker shared a personal anecdote about playing Pokémon Red on the original Game Boy in 1996 and later FireRed on the Game Boy Advance in 2004. In FireRed, a subtle change was made to the guard's drink requirement, causing an AI agent trained on Gemini 2.5 to become obsessed with finding or giving the guard tea. This example illustrates how context poisoning can occur, affecting the model's performance and behavior.