MiniCorp: The Last Mile of the AI Agent Firm

MiniCorp. The Last Mile of the AI Agent Firm. By Jingying Zeng and eight others, from Microsoft, the University of Illinois Urbana-Champaign, and three other institutions. Published on the fifth of October. No code repository is linked to it.

The authors frame the last mile toward enterprise AGI as a company that runs itself. Getting there means training and adapting agents on longitudinal enterprise data: records of a business unfolding over time. That data is scarce. It is costly to acquire, and privacy constraints often restrict it. Even where historical archives exist, they are frequently incomplete, and they have a deeper limitation. An archive records only what actually happened. It cannot show what would have followed if a different decision had been made, and those alternative outcomes are exactly what an agent learning to decide would want to see.

MiniCorp is an office simulator built to address both gaps at once. It is a place to study how agents can collectively run a company, and while they do, it generates enterprise data at scale. The demonstration is an e-commerce company. Rather than mining a past that cannot be rerun, MiniCorp produces a history that can be. The agents are not only the subject of study here. Their activity is also the raw material, so the longer they run the business, the more of the very kind of data the authors say is missing piles up.

The simulator connects two interacting worlds. The external world models customers, dynamic competitors, and market mechanisms. The internal world is made of agents that observe events, discuss their options, and make strategic decisions. Those decisions have lasting effects on the market, and the feedback the market returns informs what the firm decides later. The loop runs in both directions.

As firm and market interact, MiniCorp continuously records the agents' communications and decisions. Each record preserves the information that was available at the moment of decision, alongside the business results that followed. Then comes the piece archives cannot offer. Checkpointing lets the same situation be replayed under a different decision, so outcomes can be compared side by side. That is the counterfactual that a static archive never contains. The starting situation is held the same, and only the decision, and whatever follows from it, is allowed to differ.

Fidelity is evaluated end to end, against patterns reported in empirical studies of real markets. The authors give that check a second purpose: it provides agents with realistic market feedback, and it reduces the risk that they learn to exploit flaws in the simulator rather than learning to run a business. In the experiments, agents coordinate across roles and adapt their decisions to market feedback. One behaviour stands out. Given explicit long-term strategic guidance, the agents sustain exploration in advertising even though early returns are weak, rather than abandoning it at the first poor result.

What MiniCorp offers is two things in one environment. It is a setting for studying companies run by AI agents. And it is a scalable source of enterprise data with both a longitudinal dimension and a counterfactual one, intended for training and evaluating agents. The simulator answers the data problem by manufacturing the history, including the branches that real firms never get to see.