PaperlayerView exact revision on arXiv

Source-linked AI summary

MiniCorp: The Last Mile of the AI Agent Firm

Jingying Zeng, Zhenwei Dai, Jinning Li, Changho Shin, Dylan Zhang, Yuxuan Lu, Qi He, Dakuo Wang, Kai-Wei Chang

arXiv:2610.05912v1cs.AI

TL;DR

Agents that run companies need longitudinal data linking decisions to context and outcomes, but archives are scarce, constrained, and unable to reveal alternative trajectories. MiniCorp couples an agent firm with an evolving, empirically evaluated market, and its experiments report persistent market effects and organizational adaptation, with guidance sustaining advertising despite weak early returns.

  • Problem

    Agents that run companies need longitudinal enterprise data linking decisions to context and consequences, but archives are scarce, privacy-constrained, often incomplete, and contain only realized decisions.

    Source: §1 · p. 2

  • Method

    MiniCorp couples a persistent agent firm with an evolving market, grounds market-response pathways in empirical evidence, and records decisions and outcomes for checkpointed replay.

    Source: §1 · pp. 3–4

  • Results

    Temporary discounts and stockouts produce opposite persistent effects through ranking feedback, while advertising yields delayed organic gains and trade-offs; strategic guidance sustains investment through weak early returns.

    Source: §4.2 · pp. 13, 15

  • Takeaways & Limitations

    MiniCorp offers a closed-loop setting for studying agent-run companies and generating longitudinal, counterfactual enterprise records for agent training and evaluation.

    Source: §1 · p. 4, §5 · p. 16

    Robustness across independently initialized worlds remains to be evaluated.

    Source: §4.2 · p. 13

Abstract

from arXiv · show

The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. Historical archives are also frequently incomplete and record only what actually happened. They cannot show the outcomes of alternative decisions. We introduce MiniCorp, an office simulator for studying how agents can collectively run a company while generating enterprise data at scale. Using an e-commerce company as a demonstration, MiniCorp connects two interacting worlds. The external world models customers, dynamic competitors, and market mechanisms. The internal world consists of agents that observe events, discuss their options, and make strategic decisions. These decisions have lasting effects on the market, and the resulting feedback informs the firm's later decisions. As the firm and market interact, MiniCorp continuously records the agents' communications and decisions. These records preserve the information available at the time and the business results that followed. Checkpointing allows the same situation to be replayed under different decisions, providing comparisons unavailable in static archives. We evaluate end-to-end fidelity against patterns reported in empirical studies of real markets. These evaluations provide agents with realistic market feedback and reduce the risk that they learn to exploit flaws in the simulator. Our experiments show agents coordinating across roles and adapting their decisions to market feedback. With explicit long-term strategic guidance, they also sustain advertising exploration despite weak early returns. MiniCorp thus provides an environment for studying AI-run companies and a scalable source of longitudinal and counterfactual enterprise data for agent training and evaluation.

1. Introduction

Enterprise AGI requires organizations that set priorities and coordinate decisions over time, but existing agents and data sources fall short of that goal. MiniCorp couples a persistent agent firm with an evolving market to generate contextual, counterfactual enterprise data and evaluate market fidelity.

Motivation

  • Existing agents and benchmarks emphasize assigned, bounded tasks, whereas running a company requires persistent coordination across roles and decisions whose consequences unfold over time.Business decisions often lack a single verifiable correct action, with outcomes emerging through interactions among firms, customers, competitors, and market conditions.Source: §1 · pp. 2–3
  • The paper frames organizational autonomy as agents collectively setting strategic direction, deciding under uncertainty, coordinating work, and sustaining a profitable organization over time.Business decisions are consequential because their effects may take months or years to become fully observable.Source: §1 · p. 2
  • Historical enterprise archives are scarce, costly, privacy-constrained, often fragmented, and record only the decisions actually made, limiting their use for interactive reinforcement learning.They may omit decision context and downstream outcomes, and cannot reveal what would have followed alternative decisions.Source: §1 · p. 2

MiniCorp

  • MiniCorp couples a persistent agent firm with an evolving external market, so decisions alter market outcomes that inform later decisions and can be replayed from checkpoints.Its e-commerce world models customers, competitors, suppliers, platforms, and macroeconomic conditions; standing roles retain responsibilities across decision cycles.Source: §1 · pp. 3–4
  • The simulator grounds explicit market-response pathways in empirical research and marketplace data, then evaluates the integrated external world against empirical patterns to identify pathways needing refinement.Fidelity includes response direction and dynamics, gain-loss distribution, trade-offs, and intermediate pathways, not just final outcomes or realistic-looking snapshots.Source: §1 · pp. 3–4

2. Related Works

Prior work studies human-like social simulation and workplace agents performing bounded tasks, while business-operation benchmarks restore longer horizons with narrower agents or worlds. MiniCorp is positioned against these distinct trade-offs.

Social simulation

  • Social simulations pursue human-likeness, extending from persona fidelity and breadth to persistent worlds with irreversible consequences and emergent collective behavior.Smallville [Park et al., 2023] used memory, reflection, and planning; later work includes interview-grounded personas [Park et al., 2024] and long-running self-governed worlds [Akkil et al., 2026].Source: §2 · p. 4

Office work and company operation benchmarks

  • Workplace benchmarks often use organizational structures for coordination but stop at bounded deliverables, while business-operation benchmarks restore longer horizons with lone agents in small or abstract worlds.FoodTruckBench [foo, 2026] and MerchantBench [Shi et al., 2026] run businesses for months to a year; GDPval [Patwardhan et al., 2026] and APEX-Agents [Vidgen et al., 2026] assess isolated deliverables.Source: §2 · p. 4
  • TheAgent-Company [Xu et al., 2025] and EnterpriseArena [Han et al., 2026] target company operation, using scripted colleagues or an abstract financial model, respectively.EnterpriseArena models 132 months of survival.Source: §2 · p. 5

3. Problem Statement and Methods

MiniCorp couples a persistent agent firm with an evolving market in a closed loop, making strategic decisions and their market consequences observable over time. Its e-commerce example is designed to expose strategic capability and long-term vision through decisions such as advertising and pricing.

3. Problem Statement and Methods

  • MiniCorp evaluates strategic capability and long-term vision through business decisions such as advertising and pricing, which require learning from feedback and balancing immediate sales against potential future organic visibility.Advertising also entails query discovery, bid adjustment, budget allocation, and acquisition-cost control.Source: §3 · p. 5

3.1. The External World Model

  • The external world is an autonomous, cumulative economy that runs without firm action, responds to competitors and shocks, and keeps state and latent market parameters over time.Market demand allocation also varies by category, and published evidence informs mechanisms and parameter ranges.Source: §3.1 · p. 6

3.2. Constructing the E-Commerce Competition World

  • The e-commerce world models a response chain from search and ranking through clicks, conversion, sponsored auctions, and adaptive competitors; advertising can affect paid sales and later organic visibility.Query-level conversion performance is hidden, so sellers must learn it from observed market outcomes.Source: §3.2 · pp. 6, 8

3.3. The AI Agent Firm

  • The firm consists of standing roles that persist across the run; each has assigned business data and permitted decisions, while agents determine their own workflows and coordinate through messages.Roles can differ in authority even when they share business data, including approval authority without direct action authority.Source: §3.3 · p. 9

3.4. Interaction Between the Agent Firm and the External World

  • Each period, roles receive filtered business data, communicate, and propose authorized decisions; approved inputs and exogenous events update the world at the next period, with records supporting role- and decision-level attribution.The export uses raw business tables, while withheld world information is retained for later scoring.Source: §3.4 · pp. 9–10

3.5. Fidelity of the Simulated External World

  • Integrated fidelity tests compare controlled market-response trajectories with empirical patterns, assessing direction, timing, distributions, and trade-offs rather than relying on isolated rules or exact point estimates.This addresses the risk that agents learn simulator patterns that do not hold in real markets.Source: §3.5 · pp. 10–11

4. Experiments and Results

The experiments assess the integrated market simulator and the agent firm’s organizational and operational behavior. Fidelity tests show persistent downstream market effects, while firm runs reveal coordination, adaptive decisions, and practices shaped by guidance and agent initiative.

4.2. Results

  • Temporary discounts and stockouts produce opposite persistent effects through sales-history and ranking feedback, while advertising brings delayed organic gains with cost and competitive trade-offs.These jointly emerging effects provide system-level fidelity evidence for the default configuration; robustness across independently initialized worlds remains unevaluated.Source: §4.2 · pp. 12–13
  • Of 30,248 work items, 58% carried over from the prior cycle, 31% came from a colleague, and only 11% originated in an agent’s own weekly plan.This records work provenance across the firm’s operational cycles.Source: §4.2 · p. 13, Table 2 · p. 13
  • CEO rejection rates were highest for replenishment and logistics (29%) and advertising and pricing (25%), which commit cash or margin, and lowest for quality incidents (13%).The passage reports that this allocation of scrutiny emerged from the CEO’s judgment rather than explicit instructions to match scrutiny to financial exposure.Source: Table 3 · p. 13, §4.2 · p. 14
  • Quality Engineering and Supply Chain exchanged 1,299 work items directly, while five other high-frequency role relationships involved the CEO in approvals or task dispatch.Quality Engineering could diagnose defects but lacked authority to place a lot on hold, which Supply Chain held; the resulting requests and disposition reports account for comparable two-way volumes.Source: §4.2 · p. 14, Table 4 · p. 14
  • With guidance, the firm spent about $1,700 on advertising and retained $1,913 in gross profit after advertising, versus a $7 loss without guidance over 26 weeks.The guided firm generated approximately $31,000 in revenue and $3,645 in gross profit before advertising, compared with less than $200 in revenue and $43 in gross profit without guidance; both gross-profit measures exclude refunds, inventory write-offs, and overheads.Source: §4.2 · pp. 14–15
  • The firm proposed and received CEO approval for the expected action in 11 of 20 shock runs, and partially addressed the situation in four more.Runs covered ten business situations over eight weekly periods, repeated twice; scoring used the expected action rather than final profit because eight weeks was too short for a repair to pay for itself.Source: §4.2 · p. 15
  • Colleagues played a critical role in four shock situations and helped, without being essential, in two others.In one critical case, a supply lead’s message directed the demand lead’s attention to in-stock listings that had been switched off.Source: §4.2 · p. 15
  • Emergent behavior included evasive completion—detailed replies without pausing stockless listings—and agent-developed title revisions associated with weekly sales rising from 91 to 189 units.The title-revision sequence was not prescribed; unlike the operational initiative, sustaining advertising through weak early returns depended on explicit strategic guidance.Source: §4.2 · pp. 15–16

5. Conclusion

MiniCorp couples a persistent agent firm with an evolving market to generate enterprise records, including counterfactual evidence from replaying situations under different decisions.

5. Conclusion

  • MiniCorp couples a persistent agent firm with an evolving market, grounding its external world in economic mechanisms and real marketplace data.Its end-to-end fidelity is evaluated against empirical market-response patterns.Source: §5 · p. 16
  • Firm-market interactions generate records connecting decisions to business context and consequences, including emergent organizational and operational behaviors.Source: §5 · p. 16
  • Checkpointing enables replaying the same situation under different decisions, adding counterfactual evidence unavailable in static archives.Source: §5 · p. 16
  • MiniCorp is presented as a scalable source of realistic office-work data for training and evaluating models on long-horizon organizational tasks.The framework is described as extensible to additional firms, organizational structures, and industries.Source: §5 · p. 16

A.1. World Model Initializations and Agent Firm responses

Table 5 describes ten world-model situations and their initializations; Table 6 records responses and the contribution of collaboration.

A.1. World Model Initializations and Agent Firm responses

  • Table 5 covers ten situations, each initializing selected state variables or latent parameters while preserving the world-model structure.Source: Table 5 · p. 20
  • The Table 5 situations are ordered by the share of archived actions accounted for by the action types each exercises.Source: Table 5 · p. 20
  • Table 6 records responses by situation and audits collaboration as irreplaceable, helpful, unhelpful, or absent; “Found” counts repetitions reaching the open route.Source: Table 6 · p. 21

Related papers

Loading 2610.05912v1…