The problem is not model capability. It is context. An agent reasoning about the wrong version of a codebase produces output that is plausible but wrong in ways that are hard to catch.