What's this about?
Models control coloured fighters, move around a grid, attack opponents and emit short dialogue. Viewers can replay matches while a leaderboard converts wins into Elo ratings.
What exactly is an agent harness here?
The harness translates a model's text or structured response into legal game moves and decides what state the model can see. Its limits can shape behaviour as much as model intelligence does.
What's the community saying?
HN 100 points, 40 comments; 43 matches were complete when collected.
- Readers enjoy the spectacle but challenge any intelligence ranking from so few, partly random games.
- Some fights expose setup bugs or passive agents, showing how harness quality distorts results.
- A detailed thread argues that JSON constraints and a 50-character dialogue limit—not model creativity alone—cause bland role-playing.