Open datasets and evals for agents that act.

We build open datasets, evaluations and small models for frontier AI labs and applied teams, and we publish the research behind them. Our work is aimed at two domains: the agentic web, where we measure and train AI agents that act across the internet, and physical AI, where we start with retail: shelf data and small vision models that run inside stores.

Collecting high-quality behavioral data at scale is expensive, unglamorous, and undervalued. We think it is also the most important bottleneck in AI progress. Frontier labs can train better models if they have better data. Applied teams can ship better agents if they can measure agent behavior precisely. We exist to help close both gaps.

Our first applied arm is Machine Commerce, which covers everything in agentic commerce: products and services that help merchants get ready for AI agents that shop, compare and pay on behalf of people. Our research on how those agents behave feeds its work, and its work tells us what to measure next.

What’s next: CartBench, a benchmark for commerce agents built on CartTrace, and Commerce-1, a small model that spots when a shopping agent is about to fail. To work with us or get early access, write to founders@generalmachines.ai.