Researchers introduced MiniCorp, an office simulator for studying how AI agents can collectively run a company while generating enterprise data at scale. The paper was published on October 5 and submitted to Hugging Face by Dylan on October 7. Its authors are Jingying Zeng, Zhenwei Dai, Jinning Li, Changho Shin, Dylan Zhang, Yuxuan Lu, Qi He, Dakuo Wang and Kai-Wei Chang. They say training and adapting agents for enterprise use requires long-term business data, which is scarce, costly to obtain and often restricted by privacy rules. Historical records can also be incomplete and show only decisions that were actually made, leaving the outcomes of alternatives unknown.

MiniCorp links two simulated settings, demonstrated through an e-commerce company. Its external setting models customers, changing competitors and market mechanisms; its internal setting contains agents that observe events, discuss options and make strategic decisions. Those decisions affect the simulated market over time, and the market feedback informs later decisions. The system continuously records agent communications and choices alongside information available when decisions were made and the business results that followed.

A checkpointing feature lets researchers replay the same situation with different decisions, enabling comparisons that static archives cannot provide. The researchers evaluated whether the simulator reproduces patterns reported in empirical studies of real markets. They say this evaluation gives agents realistic market feedback and reduces the risk that agents learn to exploit flaws in the simulator. Experiments showed agents coordinating across roles and adjusting decisions in response to market feedback. With explicit long-term strategic guidance, agents also continued exploring advertising despite weak early returns. The authors present MiniCorp as an environment for studying AI-run companies and a scalable source of long-term and counterfactual enterprise data for agent training and evaluation.

Researchers introduced MiniCorp, an office simulator for studying how AI agents can collectively run a company while generating enterprise data at scale. The paper was published on October 5 and submitted to Hugging Face by Dylan on October 7. Its authors are Jingying Zeng, Zhenwei Dai, Jinning Li, Changho Shin, Dylan Zhang, Yuxuan Lu, Qi He, Dakuo Wang and Kai-Wei Chang. They say training and adapting agents for enterprise use requires long-term business data, which is scarce, costly to obtain and often restricted by privacy rules. Historical records can also be incomplete and show only decisions that were actually made, leaving the outcomes of alternatives unknown.

MiniCorp links two simulated settings, demonstrated through an e-commerce company. Its external setting models customers, changing competitors and market mechanisms; its internal setting contains agents that observe events, discuss options and make strategic decisions. Those decisions affect the simulated market over time, and the market feedback informs later decisions. The system continuously records agent communications and choices alongside information available when decisions were made and the business results that followed.

A checkpointing feature lets researchers replay the same situation with different decisions, enabling comparisons that static archives cannot provide. The researchers evaluated whether the simulator reproduces patterns reported in empirical studies of real markets. They say this evaluation gives agents realistic market feedback and reduces the risk that agents learn to exploit flaws in the simulator. Experiments showed agents coordinating across roles and adjusting decisions in response to market feedback. With explicit long-term strategic guidance, agents also continued exploring advertising despite weak early returns. The authors present MiniCorp as an environment for studying AI-run companies and a scalable source of long-term and counterfactual enterprise data for agent training and evaluation.