A new experiment using Pokémon Showdown suggests that conversation can materially improve how large language model agents coordinate, although the performance gain comes with a substantial computing cost. Researcher Oliver Grabner built a four-player version of the game's two-versus-two format in which two LLM agents share a team, exchange private messages and choose their moves independently.
The environment uses the official Pokémon Showdown simulator wrapped in a Gym-style interface. Each of four participants controls one creature from a team of three. On every turn, an agent receives a plain-language account of the battle and a numbered list of legal moves. It can send a message to its teammate or commit to an action, after which the teammate is informed of the choice.
Pokémon battles offer a useful coordination test because information is incomplete and outcomes include randomness. Opponents' move sets are hidden, teammates must divide targets, and a good plan may require one player to support or protect the other rather than simply selecting the move with the most immediate damage. The setup differs from common multi-agent workflows in which one agent hands a discrete task to the next along a fixed pipeline.
Grabner compared four communication arrangements using Qwen3.8-27B agents against a scripted opponent that selected maximum-damage moves. Teammates allowed to converse freely won 64 percent of their battles. Agents that could not talk won 42 percent. Allowing each teammate a single message produced a 52 percent win rate, while assigning one model to control both Pokémon yielded 54 percent.
The experiment used seeded battles and played every team draw from both sides so that each communication condition faced the same situations. The result supports the idea that dialogue itself contributed to the improvement: one model with visibility over both choices performed better than silent independent agents, but worse than two agents that could continue exchanging messages before committing.
Communication was expensive. Free conversation consumed about 978,000 prompt tokens per battle, 3.8 times the roughly 259,000 tokens used by the no-talk condition. Each additional message required another model call that reread the battle state and preceding discussion. That creates a direct tradeoff between coordination quality and inference cost for systems designed to solve tasks through repeated agent interaction.
Additional matchups indicated that the environment could distinguish models of different strengths. Teams using the 27-billion-parameter model won 96 percent of battles against teams of 2-billion-parameter models and 74 percent against 9-billion-parameter teams, compared with 64 percent against the scripted maximum-damage opponent.
The project remains an exploratory benchmark rather than evidence that the same gains will appear in every real-world multi-agent application. Grabner plans to separate the value of dialogue from simply seeing a partner's selected move, develop a cross-model leaderboard and investigate reinforcement-learning approaches. The current result nevertheless provides a controlled measurement of both the benefit and the cost of letting autonomous teammates talk.
