Play Pokemon Showdown doubles with two agents talking to each other

Grabner wrote a harness that lets language-model agents play battles on Pokemon Showdown, and posted the write-up to Hacker News. The interesting part is 2v2 doubles: each side is driven by agents that can send messages to each other between turns, so coordination has to be argued rather than computed. Teams came from a fixed, randomly seeded pool, and he reran every matchup with the sides swapped to cancel out team strength. He frames it as a game-theory setting with hidden information, since the opponent's sets are not visible and damage varies within a range.

In the thread he says the agents do not play the public ladder yet; he played against them himself and reports they play better than an amateur, but he was able to beat them a good amount. He gives no win rates. A commenter asked for the team selection details to be in the methodology, and Grabner agreed. Another asked whether reasoning effort was varied, which he did not answer in the thread.

Commenters pushed back on the setup: a plain decision model will score better at single-agent play, and the point here is the teammate communication it cannot do. One noted that a reinforcement-learning bot built on a rewritten battle engine already sits near the top of the leaderboard, so this is not a strength record. Another called Showdown competitive wiki reading more than a game, which Grabner half agreed with. Damage rolls are not truly random, and whether the agents account for them is unclear.