(A veiled reference to ARC-AGI with my usual cynicism).
Edit: to make this comment a bit more substantial it looks to me like Opus Magnum is a game where the player must design a custom solver for a puzzle, so it makes sense that there would be "bots" automating the process (where basically writing a bot replaces the game interface with a programming language). The challenge would be in designing a bot that can solve entire classes of Opus Magnum puzzles, maybe even the entire class. In modern AI everything has been done by learning so an "AI programmer" would typically try to machine-learn such a general solver. Unfortunately that is an open problem for modern AI, how to learn general solvers for entire classes of problems.
On the other hand, I note that the competition is based on what I understand are procedurally generated puzzles. Now, a program that can generate puzzles should also be able to solve puzzles (in the same way that an automaton that can recognise a string can generate it). Learning to identify a generator from example of its output is an open problem in AI. Learning to identify a solver from examples of its output when run as a generator is, well, even more open.
Anyway all this is too much like work for me to be a fun challenge.