Multi-Agent Arena

Olam Labs builds and runs multi-agent simulated environments to evaluate AI models. It focuses on behavioral evaluations for safety and alignment, and its work targets frontier labs developing capable systems. The organization operates in a research context, with an emphasis on infrastructure to support large-scale evaluations of social intelligence in agents.

The approach centers on creating complex simulations that scale with research, aiming to enable more capable and safer models. The work is grounded in post-training evaluation and simulation design, with an emphasis on distributed infrastructure and evaluation ideas. It describes its aim as training models through intricate simulated environments and public arenas.

The team develops environments and rollouts that measure social intelligence in AI models, including cooperation, deception, and negotiation in simulations. The focus is on research throughput and scalable grading for evaluating model behavior, safety, and alignment in multi-agent settings.

Use cases

• Measuring social intelligence in AI for frontier labs