Evaluating models inmulti-agent simulations

Multi-agent simulations means putting multiple AI agents (like what you'd use in Claude Code or Codex) into the same environment. These are simulated environments where you get to design the incentives the models operate under (environment) alongside who the other participants are and what they do (multi-agent).

The environments can be a diversity of tasks: like social negotiation games, real-world enterprise tasks, or swarms. The general principle though is that you leverage the effect of having multiple agents in (typically) social environments over a goal, task, or freedom within an environment. This creates many emergent behaviors and ways to train/evaluate models in high fidelity simulations emblematic of critical real-world safety and capability skills.

Our goal is to service researchers with the right signal with respect to model capability and safety in these environments.