AI coding agents are becoming more adept at generating code, but a significant challenge lies in assessing their engineering judgment. Simply passing tests is insufficient, as it doesn't guarantee an AI's understanding of architectural context, project constraints, or historical decisions. As AI moves from snippets to repository modifications, this contextual correctness becomes crucial. Current benchmarks like SWE-bench are evolving to address real-world software engineering issues, yet they also face challenges with task quality and contamination. The author proposes evaluating AI agents on their ability to adapt to relevant context shifts while maintaining stability when irrelevant context changes. This involves testing for context adaptation and context stability, moving beyond simple pass/fail metrics. The focus should shift from "how much context" to "which context," "when," and "how reliably." Treating repository context as a dynamic entity with a lifecycle is essential for long-lived projects. Future benchmarks should be dynamic, reflecting the evolving nature of software development. The ultimate question is whether AI agents can maintain sound engineering judgment amidst changing software environments, encompassing architecture, requirements, dependencies, and team conventions. This complex problem requires diverse thinking approaches, which can be structured using defined "lenses" for visualization, explanation, and analysis.