This cutting-edge field explores intelligent systems engineered for autonomous action, marking a significant advancement beyond current generative AI capabilities. Recent original work includes a Google Research-developed AI system that empowers scientists to craft expert-level empirical software. This system, leveraging large language models and iterative optimization, proficiently proposes, implements, and validates solutions, demonstrating expert-level performance across diverse scientific benchmarks from genomics to time-series forecasting. Its impact is profound, drastically reducing exploration time and allowing researchers to concentrate on core challenges. Simultaneously, the broader development of these autonomous systems grapples with critical scientific frontiers. These challenges encompass designing native embedding languages for effective inter-system communication, establishing secure protocols for context sharing while preserving privacy, modeling complex negotiations informed by behavioral game theory, and embedding personalized commonsense policies. Overcoming these hurdles is paramount for the safe, reliable, and effective real-world deployment of such advanced intelligence.