Researchers have developed AgentMercury, a framework for generating large-scale, realistic business environments that can be used to train AI agents. Instead of creating specific tasks and benchmarks, AgentMercury creates a persistent world with entities, services, and tools from which diverse tasks can emerge. This approach has been shown to improve the performance of reinforcement learning models on various benchmarks, including enterprise workflows and out-of-domain tasks. The framework also allows for the construction process itself to be learned, enabling AI agents to generate new environments. --- Why it matters: This matters because it provides a more realistic way to train AI agents for business scenarios, allowing them to learn from diverse tasks and interactions that emerge in complex real-world workflows. Source: https://arxiv.org/abs/2608.20634