Kaggle: Pokémon TCG AI Battle

Top 7.6% (519th of 6,807)

Kaggle: Pokémon TCG AI Battle
Munkidori
Marnie’s Grimmsnarl ex
Froslass

This summer, two of my good friends and I decided that spending some evenings and weekends on a big Kaggle competition might be a great way to assess just how much we could learn and achieve in an unfamiliar domain via well-driven, agent-primary R&D. We all had full-time jobs, no formal ML experience, and zero Kaggle know-how, but did have a custom multiplexer/runtime that supports realtime collaboration between multiple human-agent pairs.

Doing sustained work in such an organizational setup, where 1) individual human-agent pairs work nonstop on distinct but interconnected parts of one shared project, and 2) everyone can ping anyone else’s agent at any time and ask questions, submit requests, or offer steering suggestions, was certainly an experience — perhaps a glimpse of the future of work, at least in certain environments. The latter was particularly interesting, in how this kind of direct access seemed to simultaneously alleviate and exacerbate the now-commonly recognized “meat proxy” issue.

While our final placing wasn’t satisfactory (albeit perhaps still respectable considering the team’s raw-newbies makeup), I felt that we suffered mostly from a few key strategic miscalls rooted in our lack of domain experience, not in the way we chose to work. We’ll give this method another try in the coming months, and I’m sure I’ll have deeper thoughts to share then.

Team
DittleDev
Period
Jul–Aug 2026
Card art
© Pokémon / Nintendo / Creatures / GAME FREAK