AgentMercury Synthesizes Business Worlds to Train Agents

Researchers Minbyul Jeong and Chanwoong Yoon have introduced AgentMercury, a scalable framework designed to synthesize executable environments from high-level business scenarios. Traditional AI agent training relies on manually constructed, task-specific environments that struggle to replicate the complexity of real-world workflows. AgentMercury shifts this paradigm by instantiating persistent virtual worlds complete with distinct entities, services, tools, operational states, and executable cross-service rules, allowing diverse tasks to emerge naturally.

Using this framework, the researchers constructed 4,783 executable environments spanning 14 industries and 50 countries. These environments were then used as training substrates for reinforcement learning. Despite not targeting specific evaluation benchmarks during the generation process, policies trained on these business-oriented environments showed substantial improvements across both enterprise workflows and out-of-domain benchmarks, including reasoning, coding, scientific computing, and tool use.

The empirical results highlight the effectiveness of this training method. For instance, the Qwen3.5-4B model improved its score from 12.3 to 15.7 on the EnterpriseOps-GYM benchmark and saw its performance rise from 45.9 to 56.0 on the AIME26 benchmark. Furthermore, the researchers demonstrated that the environment construction process itself is a learnable skill. By fine-tuning the Qwen3.5-35B-A3B model on construction traces, they increased its success rate in authoring executable worlds from a mere 3.3% to 83.3% on held-out business scenarios.

For AI practitioners, AgentMercury represents a major shift in how agents are trained and deployed. Instead of spending hours manually coding brittle, task-specific training simulations, developers can now use large language models to generate diverse, verifiable business environments at scale. This approach not only reduces the engineering overhead of building training pipelines but also produces agents with highly generalizable problem-solving abilities that translate well to real-world enterprise operations.