Humans still needed: Learnings from refactoring an ancient codebase

AI agents have become fantastic at green field projects. However, getting them to work correctly on existing, complex codebases can be a challenge. I recently ported a 100k loc Obj-C/OpenGL iOS game from 2011 to modern iOS with the help of Opus 4.6.

This required much more than a simple prompt. The agent lacked the creativity to solve this task, it was often overwhelmed and some problems required a change of abstraction.

The talk will go through the initial setup, how I improved it during the project, what the agent struggled with, and how I was able to guide it.