Multi-Agent Pokemon Showdown
Oliver Grabner's Pokemon Showdown environment for studying private communication and coordination between two LLM teammates.
What happened lately
Multi-Agent Pokemon Showdown timeline
This community research project wraps Pokemon Showdown Multi Battles to compare two agents talking freely, exchanging one message each, acting without conversation, or a single model controlling both teammates. Each player controls one active Pokemon from a team of three. Fixed seeds and mirrored team assignments support comparisons across conditions.
Grabner’s October 10 report describes Qwen3.8-27B experiments against a scripted maximum-damage bot and smaller-model teams. The reported results are author measurements, not AITrending reproductions or human ladder evaluations. The project was shared on Hacker News on October 22; that discussion date is separate from the report’s original publication date, whose time of day is unspecified.
The framework needs Python 3.11 or later and Node.js 18 or later. Its English README documents simulator setup, replays and model-backed batches with an OpenAI-compatible client. Model serving or API access must be supplied separately. A cross-model leaderboard and reinforcement-learning training remain proposed follow-up work.