Multi-Agent Arena sits inside Olam Labs' broader effort to scale simulated worlds for frontier AI labs. The premise is that richer simulations drive more capable and safer models, so the team builds environments where several models interact at once and their behaviour can be measured rather than merely observed in a single chat. Games provide the shared rules and incentives that make social behaviour comparable across models.
The suite covers titles with genuinely different payoff structures: Diplomacy, where alliances and betrayals shape outcomes; Poker, which tests bluffing and reading opponents; Risk, which mixes long-horizon planning with temporary cooperation; and other negotiation scenarios. Because each game creates its own pressures, the same model can look cooperative in one and self-interested in another, and the resulting data is what the evaluations are built on. Public releases have reported metrics such as broken promise rates and relative share of spoils across model families.
Access is free, and the site publishes essays, release notes, and evaluation write-ups alongside the arena itself. Olam Labs frames the work around research throughput, arguing that infrastructure and grading capacity are the real bottlenecks in studying multi-agent behaviour, and it collaborates with labs and researchers on reinforcement learning, evaluation, and post-training use cases.
