> As of today, none of the existing models can meaningfully handle mid-size tasks on five services with 10k+ LOC each

My FAANG's codebase is a few orders of magnitude larger and agents do an excellent job of handling mid sized tasks completely autonomously.