Startup Plurai adapts autonomous vehicle simulation techniques to improve AI agent reliability

Plurai, a new startup, is applying simulation methodologies, similar to those used in developing self-driving cars, to enhance the reliability of AI agents. The company, which recently secured $13 million in seed funding led by Team8, aims to address the current limitations of agents, which often require close human oversight. Initial benchmarks show models like Claude Fable 5.2 achieving 82.9% success on tasks, while EON's testing found Claude Opus 4.8 at 76.8% accuracy in a simulated corporate setting. Co-founders Ilan Kadar and Elad Levi previously worked with autonomous driving technology firms, including Waymo. Kadar noted that while building AI demos is easy, achieving reliable and cost-effective systems is difficult. Plurai builds digital twins of a company's internal systems to map all possible agent flows, allowing for pre-training and fixing of edge cases. Furthermore, Kadar stated that while powerful general models are impressive, they lack specific business knowledge. Plurai's unique value proposition involves using gathered data to post-train smaller, open-weight models tailored to specific clients. Tests with Nvidia and Nebius demonstrated these models could outperform mid-tier frontier models at a fraction of the cost. Reah Miyara, a Google DeepMind researcher and angel investor in Plurai, highlighted that the company’s strength lies in generating synthetic data to tune models, addressing IT leaders' concerns about the unsustainable cost of frontier models.