Presentation: The Right 300 Tokens Beat 100k Noisy Ones: The Architecture of Context Engineering

Baruch Sadogursky and Patrick Debois discussed why coding agents fail due to bloated context windows and stuffed prompts. They explained practical context engineering fixes, such as lazy-loaded skills, versioned context artifacts, externalized memory banks, and LLM-as-a-judge evaluations. The speakers emphasized that software architects and engineering leaders can turn raw markdown files into reliable agentic workflows.

Patrick Debois, a practitioner and researcher exploring how AI agents reshape software development, and Baruch Sadogursky, a DevRel and Context Engineering Management at Tessl AI, shared their insights on the architecture of context engineering.

Sadogursky demonstrated how even a small project with extensive context can lead to errors. Despite providing ample context, one of his endpoints encountered an AssertionError, expecting a 404 response for a missing order but receiving a 500 response instead. Debois suggested that the issue might stem from stuffing too much information into the context window.

The duo highlighted that coding agents, like Claude, have system prompts and chat conversations that contribute to the context window, eventually causing it to exceed its capacity. They emphasized that the system is stateless, and every interaction, including past logs and chats, adds to the context window, leading to inefficient agent behavior.

The speakers explained that context engineering is distinct from prompt engineering, with the latter focusing on crafting perfect queries while the former deals with managing and optimizing the overall context. They stressed that stuffing too much information into the context window leads to inefficiencies and unintended behavior in coding agents.

By adopting strategies like lazy-loaded skills, versioned context artifacts, and externalized memory banks, engineers can effectively manage context and enhance the performance of AI coding agents.