The interesting part of the article is the cite: <a href="https://www.researchgate.net/publication/318811113_Measuring_Program_Comprehension_A_Large-Scale_Field_Study_with_Professionals" rel="nofollow">https://www.researchgate.net/publication/318811113_Measuring...</a><p>A revelation struck me as I read this: if software development consists of 5-10% programming or editing (as evidenced by the study and confirmed by my own experience) and over 50% comprehension, then the current state of AI coding agents is quite poor. This is because they only understand a very small portion of your software, and this portion is a very limited context that they can compare with what they already know. However, there is hope; MCP will expand this context, the financial cost and what you think about data privacy when you give AI access to anything related to a project (like mail, confluence, meeting notes and git history).