Beyond the Context Window - The Architecture of a Trustworthy AI Co-Designer

This article discusses the challenges of building Nova, a persistent AI co-designer for board game designers that works across months of sessions. It argues that the key bottleneck is not context window capacity but the architecture around it: deciding what to leave out, ensuring work survives interruptions, monitoring costs, and maintaining correctness as agent output decays. The author presents four disciplines: cutting the leash (avoiding unnecessary context), sending less data, verifying outputs even with partial context, and metering hidden costs. It is a technical field report on production issues and solutions for long-running AI collaborators.