MiniCorp: The Last Mile of the AI Agent Firm

The last mile toward enterprise AGI is a company that runs itself, with autonomous digital workers performing real jobs. However, the enterprise data required to train or adapt these workers is locked behind privacy barriers, either sold at a premium from the archives of bankrupt firms or left unusable until expensive experts label it. In this work, we introduce MiniCorp, an office simulator that provides a live environment for studying autonomous organizations while generating grounded enterprise data at scale. Using an e-commerce company as our demonstration, we simulate two interacting worlds: (i) an external market that runs on its own and emits events, grounded in Capitalism II, a mature commercial business-simulation engine, and calibrated against real markets by a deep research engine rather than being hand-authored; and (ii) an internal agent firm of five standing roles that observe those events, deliberate, and act back on the market, with every action advancing a shared clock from which the external world re-simulates the next week of events. MiniCorp is an agentic environment that enables scalable and efficient enterprise data generation, addressing the data scarcity problem. The company runs continuously and leaves behind the records that a real company would, such as messages, emails, tickets, and documents, at a volume and speed that no privacy-constrained corpus can match. Every artifact has full provenance linking it to the market state and the decision that produced it. Because its state is checkpointable, the same situation can be replayed under different decisions, yielding counterfactual pairs that a static archive cannot supply.