Andon Labs studies AI safety by letting agents run real businesses, using live deployments and simulations to expose failures before autonomous organizations become widespread.