Presentation: The Right 300 Tokens Beat 100k Noisy Ones: The Architecture of Context Engineering
| Source: InfoQ AI/ML
Tags: context engineering, Claude Code, coding agents, LLM, Tessl AI, InfoQ
Engineers at QCon London argued that most coding agent failures stem from bloated context windows, not model capability — and demonstrated that lazy-loaded skills, versioned context artifacts, and external memory banks outperform stuffing everything into a CLAUDE.md file.
Details
Baruch Sadogursky (Tessl AI) and Patrick Debois presented at QCon London on context engineering antipatterns in coding agents. Their core demonstration: loading a CLAUDE.md file with all project documentation does not reliably help agents — it can hurt by burying relevant information in noise or running the model past its effective context window. Their live example showed Claude adding error handling that broke a 404 test (returning 500 instead) after receiving a maximally loaded context. Their proposed architecture centers on right-sizing context at the right time. Key patterns: lazy-loaded skills (instructions loaded only when relevant to the current task, not upfront), versioned context artifacts (treating prompt context like a software dependency — traceable, updatable), external memory banks (storing session state outside the context window), and LLM-as-a-judge evaluations to measure context quality rather than just task outcomes. The presentation is from a conference talk, not a published paper, so the content available is a transcript excerpt. The techniques described align with broader trends in agent architecture: retrieval-augmented context, just-in-time instruction loading, and structured memory separation. Directly applicable to teams building or maintaining coding agents at scale.