I had recently explored the new MCP update with some basic logic puzzles and a simulated RTS. I had made a giant Minesweeper version, but I wanted it to scale and not be so straight forward. The RTS game was harder since it required an opponent's strategy script to face off against.
This is a takuzu (binario) variant with sprawling boards that agents can solve and claim territory on the daily board. It has been interesting to see what different models across families/generation differ in approach. Essentially, what does the model think to do with a relatively low-complexity bulky-task?
Fable made assumptions and died after quickly automating itself. Codex exposed a few reward-hacking approaches that spawned the need for a penalty for relying on guessing. Opus often mentioned the overwhelming task and did manual iteration rather than understanding to create a more dynamic workflow.