There are two lines of work behind this. One is Meta's CICERO, which played Diplomacy at a human level. The whole game was talking, cutting deals, and deciding who to trust. The other is AI that plays board games well, where Catanatron is a strong, fast Catan engine that a lot of research already uses as a backend. The negotiation space is studied, the board (or as we call it the action) space is studied. but true systems are both. Agents taking actions while ALSO negotiating.
When several language models play a full game of Catan and can also talk to each other, two things run at once. There is the action space, meaning every legal build, trade, and robber move the rules allow. There is also the negotiation space, meaning anything a player might say to persuade, threaten, or team up with someone. They feed each other. A promise only means something because of the move it changes, and a move only makes sense once you know what was said around it. Getting four agents to handle both at the same time is where the difficult and interesting behavior shows up.
Here is what we built. The game runs on Catanatron, so the rules and the board are taken care of. Each model gets a view of the state, the list of moves that are legal right now, and a short summary of what changed since it last acted, then it returns a move. Since it can only choose from the legal list, it cannot break the rules. On its own turn a model can open a negotiation round, a group chat that every other player can read and answer, with a limit on how many messages each one can send so the round ends and play goes on. Talking does not move any resources by itself. It changes which move a player picks. We log all of the moves, the messages, and each model's private reasoning, so we the players go back and check who promised what and who actually followed through.
The output is a working system and a first result: does letting the models talk change how games play out, and do they keep the deals they make. We release the code and the negotiation data. After Catan we point the same setup at bigger games like Twilight Imperium, where the negotiation even more weight.