My bot with cell wasn't very effective, and was getting complicated to modify, so I started anew with a clean, modular architecture and a way of combining different strategies, each of which is pretty simple. That gets me an effecctiv enough bot with room for improvement - most of the ants just scatter by avoiding walking on frequently-visited squares, and those that are near food and hives attack.
Now that I have that, I can start adding some more complicated pathfinding strategies, some optimisation, etc. - macro cells will probably be useful again, for optimisation and high-level pathfinding. But now I can fit them into my framework without my code getting all ugly and unwieldy.
(I'm still somewhat annoyed about how my first bot, which just goes straight and turns 5% of the time or when it hits a wall, keeps outperforming smarter attempts. Well, at least now I'm regularly beating it.)
Aichallenge.org has started their third AI contest this year: Ants.
I mentioned this in the open thread, and there was a discussion about possibly making one or more "official" LessWrong teams. D_Alex has offered a motivational prize. If this interests you, please discuss in the comments!