Most GenAI systems are easy to build and hard to change. Once in production, small updates - prompt changes, new tools, model swaps, data source updates - can introduce unexpected regressions. Hidden coupling across components, brittle orchestration, and unclear interfaces make systems increasingly fragile as they grow. This session focuses on the architectural failure modes that cause GenAI systems to calcify over time. We’ll examine how teams design for change in production by introducing clear boundaries, versioned components, and modular execution paths. The emphasis is on patterns that support safe iteration, controlled deployment, and long-lived systems rather than one-off prototypes.