Pokémon AI Agents battle with free LLMs

This is a solid proof-of-concept for agent tooling in a domain where state is compact, decisions are discrete, and failure is easy to observe. The architecture is more interesting than the game itself: it shows how far you can get with model routing, battle-state serialization, and strict action schemas before you need heavier planning.

• –The free-tier strategy matters more than the Pokémon theme: routing through LiteLLM makes multi-provider experimentation practical without forcing inference spend.
• –Full turn-by-turn observability via Langfuse is the right move here, because agent mistakes in games are only useful if you can trace why they happened.
• –The biggest technical limitation is likely state abstraction, not raw model quality; Pokémon battles reward hidden information management, long-horizon planning, and opponent modeling.
• –Human-vs-AI and AI-vs-AI modes make this useful as a sandbox for comparing prompts, providers, and tool schemas under the same environment.
• –The random fallback on invalid actions is pragmatic, but it also signals where the system still needs stronger action validation or search.