Oliver Grabner

Oliver Grabner’s project is a research environment built on Pokémon Showdown’s 2v2 format, where two LLM agents share one side in four-player battles. Rather than studying pipeline-style workflows where agents pass tasks along, it isolates the question of whether agents can communicate, negotiate, and coordinate as peers under competitive pressure.

The core feature is a pre-turn chat channel: before every turn, the two teammates can exchange natural-language messages about threats, type matchups, and intended moves, then lock in their choices. The published excerpts show agents correcting each other’s assumptions and splitting targets, which makes the coordination behavior directly observable rather than inferred.

The setup is designed as a measurement tool, tracking whether talking actually wins more battles and what that communication costs in tokens or latency. The code is open on GitHub, so the environment can be reproduced and extended.

It is aimed at AI researchers, multi-agent system developers, and anyone benchmarking LLM collaboration or negotiation abilities. Those looking for a playable game or a consumer chatbot should look elsewhere.

To get value from it, read the sample battle logs first to see the coordination patterns, then run the environment with your own models to compare talkative versus silent teams. It is a narrow but well-defined benchmark for a question that matters beyond games.