Pokemon Showdown agents battle autonomously

The interesting part here is not that an LLM can pick a move, but that the game is being turned into a structured decision problem with full state exposure each turn. That’s the right shape for cheap models, though the real ceiling will come from better state abstraction and opponent modeling, not just more providers.

• –LiteLLM is a practical way to keep the system provider-agnostic while swapping among free-tier models
• –Full-state prompts plus tool calls should outperform naive chat loops, especially when switching and weather/field effects matter
• –Langfuse visibility is a real strength: you can debug bad turns instead of guessing why the agent blundered
• –The main next step is tighter reasoning scaffolding for complex board states, like explicit threat scoring, move ranking, and self-play evaluation
• –This feels more like an agent evaluation harness than a game bot, which makes it broadly useful to AI builders