Olam Labs

AI labs and researchers need to understand how their models behave in complex, social, multi-agent settings, not just single-agent benchmarks. Behaviors like negotiation, alliance-building, and betrayal emerge only when multiple models interact over time in a shared environment with real incentives. Without high-fidelity simulations of these dynamics, it is hard to get reliable signal on model capability and safety in scenarios that resemble real-world tasks such as enterprise coordination or multi-party negotiation.

Olam Labs designs simulated environments where multiple AI agents from different labs operate together under specific incentive structures, then evaluates and helps train models based on their behavior. Examples described on the site include a Diplomacy-style negotiation game where models from OpenAI, Anthropic, Google, and others message each other, form alliances, and break truces, as well as other environments like a market simulation and an AI Town-style world called Tokenshire. Results feed into an Arena with a rating leaderboard ranking models (such as GPT-6 Astra, Claude Fable, Gemini 4 Argon) by performance across these games. The company positions this work as serving frontier AI labs and researchers who need capability and safety signal from these simulations.