Olam Labs
Olam Labs builds multi-agent simulation environments and evaluation infrastructure for frontier AI labs, enabling researchers to test model capability and safety through dynamic, competitive scenarios rather than static benchmarks. The platform powers Multi-Agent Arena, which ranks models like GPT, Claude, and Gemini across games and real-world enterprise simulations, and supports reinforcement learning workflows with robust grading designed to resist reward hacking.
- Artificial Intelligence
- Developer Tools
- Data & Analytics
- Software Only
Funding
Raised to date
$500KRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Founder details are not available yet.
Product
Problem
Researchers developing and evaluating frontier AI models lack standardized, scalable environments for assessing model capability and safety in competitive, multi-agent contexts. Traditional benchmarks are static and fail to capture emergent behaviors and strategic interactions between different AI systems, limiting the reliability and depth of performance signals for model development.
Solution
Olam Labs operates multi-agent simulation environments that generate performance benchmarks for frontier AI models. The platform provides researchers with ranked evaluations of model capability and safety across diverse simulated scenarios, using structured leaderboards to compare models like GPT, Claude, Gemini, and others in competitive agentic contexts. These environments are designed to service researchers with the right signal with respect to model capability and safety, offering a more dynamic assessment approach than static benchmarks. The platform supports reinforcement learning research and model assessment workflows, enabling iterative comparison and tracking of model performance over time. The company also builds custom RL and evaluation environments that emulate real-world complexity, including enterprise simulations, game development, and computer use scenarios, with careful attention to reward hacking and eval awareness in grading.
Target Audience
Primary users are AI researchers, reinforcement learning practitioners, and model evaluation teams at frontier AI labs and research institutions who need robust competitive benchmarks for assessing model capability and safety.
Features
- Structured leaderboards ranking models by overall performance scores across multi-agent simulation scenarios, including the public Multi-Agent Arena
- Diverse research environments that test models in competitive agentic contexts with various model families including GPT, Claude, Gemini, Grok, Kimi, Muse, GLM, and DeepSeek
- Pre-deployment safety and capability evaluations through anonymized checkpoints on Multi-Agent Arena and internal environments emulating real-world alignment scenarios
- Complex simulations of real-world settings like enterprises with many services, game development, and computer use, with care for reward hacking and eval awareness in grading
- Alignment environments across multi-agent social games and real-world enterprises that test and train models for alignment
- Support for reinforcement learning research workflows with quantitative benchmark outputs, including environments for games like Minecraft, Mindustry, and Diplomacy