AI Architecture Updates: August 15, 2026

1. Debois and Sadogursky Treat Context as an Architectural Concern for Coding Agents

Patrick Debois and Baruch Sadogursky. In a presentation published by InfoQ, Patrick Debois and Baruch Sadogursky argue that AI coding agents fail when context windows fill with unfocused information, and that context should be designed as an architectural concern rather than assembled from dumped documentation. They propose four patterns: lazy-loaded skills that activate only when relevant, versioned context artifacts that bundle documentation, rules, and best practices, externalized memory banks that persist decisions across sessions instead of relying on opaque model memory, and LLM-as-a-judge evaluations to measure whether context changes actually improve agent output. They report a demonstration in which correctness rose from 35 percent to 98 percent using structured context artifacts versus raw documentation retrieval. Source

2. LLM-Generated GraphQL Mocks Diverge at Airbnb and Expedia While the Spec Lags

InfoQ. InfoQ’s Steef-Jan Wiggers reports that Expedia and Airbnb have independently built LLM-based GraphQL mocking solutions with incompatible architectures. Expedia’s mockql-rs runs as a middleware process that annotates fields with @mock directives, while Airbnb’s approach operates at build time to generate JSON fixtures, and the GraphQL Foundation’s RFC proposes operation-level mocks stored in reserved directories but remains at Stage 0, described as a strawman with no champion listed. The gap shows how design patterns for AI-generated test data are emerging faster than the standards meant to make them interoperable, leaving teams uncertain which implementation to standardize on. Source