Artificial Intelligence · arXiv:2610.05912 v2 · submitted 5 Oct 2026
MiniCorp: The Last Mile to an AI Agent Corporation
The paper says
“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.”
more of the abstract
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.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 6 Oct 2026 · of version 1 of the paper; version 2 is newer · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A well-built enterprise-agent simulator whose architecture is genuinely novel, and whose main vulnerability — fidelity evidence from one checkpoint with no variance — is the gap the authors themselves flag; it deserves referee time and a code release. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the references say.
The paper says, at Table 1, first row (§4.2 RQ1)
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.89):
“Robustness across independently initialized worlds remains to be evaluated.”
Pith says: Objection 1 of 4 · major
Fidelity is load-bearing for the central claim: the paper states at b49 that the usefulness of the generated enterprise records depends critically on the fidelity of the external world. Yet the four fidelity tests in Table 1 are controlled interventions run from a single checkpoint, where applicable with the same random draws, and no variance is reported.
more of objection 1
This matters structurally because §3.1 samples latent preference weights (Lambda) at world creation, §3.2(IV) samples unobserved query–category conversion multipliers at world creation, and the 26-week warm-up (§3.2(II)) generates a unique incumbent sales and review history. A single checkpoint is therefore one draw from a large distribution of worlds. The authors acknowledge the gap in b272 ('Robustness across independently initialized worlds remains to be evaluated'), but the acknowledgment does not resolve the consequence for the central claim: until the four patterns are shown stable in direction, temporal ordering, and rough magnitude across multiple independent world draws, the statement that agents 'receive realistic market feedback' is supported for one trajectory only. I recommend either (a) adding robustness runs across independently initialized worlds that report the sign, ordering, and magnitude ranges of the four patterns, or (b) scoping the fidelity conclusion to the configuration tested and moving the generalizability caveat from b272 into the abstract and conclusion.
The paper says, at Place not named
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The price-response results report arc elasticities of 2.9 and 3.2 as point estimates with no uncertainty, and the claim that these are 'close to the empirical benchmark' is unquantified. The cited source (Bijmolt et al., 2005) is a meta-analysis reporting a distribution of price elasticities (median near −2.6), so whether 2.9 and 3.2 are 'close' depends on the reference distribution and on the dispersion across the weeks, query categories, or world draws over which the elasticities were computed.
more of objection 2
Please report the number of independent measurements used to compute each elasticity, their dispersion, and a defined criterion for closeness, or soften the wording to acknowledge that the comparison is qualitative.
The paper says, at Place not named
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The fidelity section describes an iterative process in which the evidence from these evaluations is used 'to identify missing or incorrect response pathways that require further refinement,' and the paper's defense against circularity is that the duration, magnitude, and distribution of the downstream effects 'were not direct calibration targets' (§4.2 RQ1).
more of objection 3
That mitigation is meaningful but not independently checkable: since neither the simulator code nor the checkpointed data are released, a reader cannot verify that the four reported patterns were not, directly or indirectly, tuned into the parameter ranges during refinement, nor audit the pathway-tracing claims. I request release of the simulator code, the checkpoint(s) used for the four tests, and the exact prompt and parameter configuration at minimum as supplementary material, so that the validation protocol can be independently re-run and separated from the refinement process.
The paper says, at Place not named
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
The guided-versus-unguided comparison appears to be a single run per condition: the text describes 'one run' with explicit long-term strategic guidance and 'the other run' without it. The causal claim that 'explicit long-term strategic guidance changed the firm's willingness to sacrifice short-term profit' rests on n=1 per condition, and LLM-driven agent runs exhibit substantial sampling variance in planning and execution.
more of objection 4
The dramatic profit gap ($1,913 gross profit after advertising versus a $7 loss) should be accompanied by at least 3–5 independent runs per condition, reporting the mean and range, to rule out that the difference reflects prompt or sampling variation rather than the guidance manipulation. This is secondary to the fidelity claim but is used to support the long-term-vision capability dimension of the framework.
6 minor objections
The paper says, at Place not named
Pith says: minor objection 1 of 6
There are typographical errors: 'stratigic capability' (§3.2), 'commucations' (§3.3), and 'orangic search rankings' (§4.2 RQ5) should read 'strategic', 'communications', and 'organic', respectively.
The paper says, at Place not named
Pith says: minor objection 2 of 6
The 'Share' column of Table 5 is obscure: the header ('share of archived actions submitted to the world model accounted for by the action types each one exercises') is not defined in the text, and the listed percentages (53.1%, 12.0%, ..., 0.2%) do not obviously relate to the ten situations or to the evaluation in Table 6. Please clarify the definition and its role in the shock-response analysis.
show 4 more
The paper says, at Place not named
Pith says: minor objection 3 of 6
The claim that CEO rejection rates scale with financial exposure ('the categories the CEO rejects most often are the ones that spend money') is supported only by visual inspection of aggregate category-level rates; a simple statistical comparison across categories, with sampling uncertainty given the differing item counts, would substantiate the stronger claim that 'the allocation emerges from its own judgment' rather than from chance variation.
The paper says, at Place not named
Pith says: minor objection 4 of 6
The advertising results report $0.53 contribution per advertising dollar and 84% competitor displacement as single-window point estimates, with the 23–32% unit increase given as a range. Please state the observation-window length for the dollar-return and displacement figures and whether the range reflects weekly variation, so the trade-off claim can be assessed.
The paper says, at Place not named
Pith says: minor objection 5 of 6
The experimental setup names the model (GPT-5.6-Sol) but gives no decoding parameters, per-agent budget or token limits, or compute cost. Since the paper claims a 'scalable and efficient enterprise data engine', reporting these details (or citing a cost table) would support the throughput and cost claims.
The paper says, at Place not named
Pith says: minor objection 6 of 6
Figure 1 and Figure 2 both have captions, but the body text does not consistently tie each figure reference to its caption, and the captions are duplicative in places. Please verify all figure cross-references.
What would settle it
Run the four Table 1 interventions on at least 20 independently initialized worlds, drawing Λ (category preference weights), conversion multipliers, and warm-up histories from the ranges specified in §3.1–§3.2, with paired same-checkpoint controls and matched random draws within each pair. Report across-world sign agreement, temporal ordering, and distribution widths, such as the interquartile range of arc elasticities and persistence shares. Pre-register a pass rule before running, for example all four response directions and pathway orderings holding in at least 18 of 20 worlds and simulated distributions overlapping the cited empirical benchmarks; if not, the fidelity evidence is initialization-specific and does not support the training and evaluation claim.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 6 Oct 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
MiniCorp is a genuine contribution to the enterprise-agent simulation line, and its soft spot is one the authors themselves flag: the fidelity evidence is four interventions from a single checkpoint, with no variance reported, so generality across the distribution of worlds is asserted but not shown.
What is actually new is the combination. A persistent multi-role firm with standing roles and no central scheduler, coupled to a self-advancing market with explicit mechanisms — multi-slot second-price auction, organic ranking, query matching, click and conversion — plus checkpointed counterfactual replay. That constellation is not in MerchantBench (lone operator on a hand-authored world), TheAgentCompany (scripted colleagues), or EnterpriseArena (abstract financial model). The architecture is also better thought through than most of this genre: export masks that drop latent parameters and fail the whole export if anything protected survives, roles that differ in authority rather than knowledge, and provenance that attributes an outcome to a role and a decision. That design work is real.
The paper is honest, which earns credit it should keep. b272 states robustness across independently initialized worlds is unevaluated; b311 discloses that gross profit excludes refunds, inventory write-offs, and overhead; b49 says flat out that usefulness depends on external-world fidelity. The fidelity tests are controlled interventions with pathway tracing, and the arc elasticities (2.9 and 3.2) sit near the empirical benchmark. The organizational results are mostly non-circular: the CEO rejection rate tracks financial exposure with no instruction to do that, and the evasive-behavior and title-editing episodes are emergent, not prompted.
The soft spots are proportionate to how the paper frames itself. The fidelity load is four patterns from one checkpoint, and the world draws latent preference weights, unobserved conversion multipliers, and a 26-week warm-up history per initialization — so one checkpoint is one point in a large distribution. Direction and temporal ordering of those four patterns could plausibly hold across draws, but nothing here shows it. That is the central gap, and the stress-test note is right to put it there. Two smaller things: no code or data is released, which means the Section 3.5 iterative refinement cannot be audited against the later-reported validation patterns (a real concern, though unprovable either way from the text); and RQ3 reads like a single run.
This deserves a serious referee, not a desk reject. The fix list is short and addressable: report variance across independently initialized worlds, ship the code and checkpoints. I would take it to reading group; a paper this honest about its own load-bearing assumption is good discussion material.
The strongest objection
MiniCorp's central claim—that its records can train and evaluate long-horizon organizational agents—rests on the fidelity of the external world (b49). The empirical support is four controlled interventions in Table 1, all run from one checkpoint and, where applicable, with the same random draws, with no variance reported. This matters structurally: §3.1 samples latent preference weights per world, §3.2(IV) samples unobserved query-category conversion multipliers at world creation, and a 26-week warm-up generates a unique incumbent history, so a single checkpoint is one point in a large distribution of worlds. The paper itself states in b272 that robustness across independently initialized worlds remains to be evaluated. If the direction and temporal ordering of the four patterns are not stable across that distribution, agents trained in MiniCorp can learn regularities that hold only in a particular draw, exactly the simulator overfitting the authors aim to avoid. The absence of released code/data also prevents auditing how much the iterative refinement described in §3.5 consulted the same patterns later reported as validation. The concern is not the direction of any single test; it is the unestablished generality of the fidelity evidence across the world distribution the environment is meant to sample.
The reader's weakest assumption identifies the same premise: the fidelity evidence is too thin to carry the claim that agents receive realistic market feedback. I agree and sharpen it: the four tests sample one initialization from a parameterized stochastic world family, so the open question is not merely 'more seeds' but whether the integrated response surface is stable across the distribution of worlds agents will encounter. The authors honestly flag this at b272 and are careful not to claim robustness they have not measured, which is to their credit. I do not make the absence of released code the primary attack, because reproducibility is an engineering artifact gap rather than a falsification of the environment, though it compounds the difficulty of checking whether the refinement loop in §3.5 was independent of the validation patterns. The RQ3 agent results are also single-run and use gross-profit measures excluding refunds, inventory write-offs, and overheads (b311), so they are illustrative rather than the main support. Since the reader already assigned CONDITIONAL and the concern confirms that assessment, the verdict should remain unchanged pending the multi-initialization fidelity check.
The referee's report
If the fidelity claims hold, MiniCorp occupies a genuinely useful niche: unlike static archives (e.g., Enron) or task-bounded benchmarks, it produces longitudinal records with full provenance linking decisions to later market outcomes, and its checkpoint mechanism yields counterfactual pairs that historical data cannot provide. The explicit modeling of intermediate pathways (search, ranking, click, conversion, auction, organic feedback) is a strength over abstract financial engines, and the validation methodology, pattern-oriented and pathway-level with magnitudes stated to be outside the calibration targets, is methodologically sound and yields falsifiable predictions (e.g., arc elasticities in the 2.9–3.2 range; 59% of the stockout shortfall occurring after restocking). Honest reporting of partial failures (Tables 5–6) and a precise definition of emergence (behavior unspecified and not a model default, b338–b339) strengthen credibility. The principal risk is evidentiary: the four fidelity tests come from a single world draw with no reported variance, and the paper itself states that robustness across independent initializations remains to be evaluated (b272).
Pith's reading of the paper (a machine's summary, not the authors')
This paper claims that the data needed to train and evaluate agents capable of running a company—longitudinal enterprise records linking decisions to their consequences—can be generated synthetically, at scale, by a closed-loop office simulator rather than harvested from scarce, privacy-restricted, frozen archives. The demonstration, MiniCorp, couples a persistent AI agent firm with an independently evolving e-commerce market, so each approved decision alters the market and the resulting feedback shapes the next round of decisions. Every message, query, proposal, and ruling is recorded with full provenance, and checkpointing lets the same state be replayed under alternative decisions, producing counterfactual pairs that static archives cannot supply. If the simulated market responses are faithful, MiniCorp is both an environment for studying AI-run companies and a scalable data source for long-horizon organizational agents.
What carries the argument. The load-bearing object is the closed-loop world transition $W_{t+1} \sim P_\Lambda(\cdot \mid W_t, D_t, Z_t)$: a cumulative external state updated by the agent firm's decisions $D_t$ and exogenous shocks $Z_t$ under latent structural parameters $\Lambda$ that the firm never observes and must infer from business data alone. Around this transition the paper builds an explicit response chain—query demand through matching, organic ranking, clicks, conversions, sponsored auctions, and competitor reactions—so that no business action maps directly to an outcome without passing through intermediate mechanisms, which is what lets agents diagnose where in the funnel they are losing value. Fidelity is assessed by pattern-oriented modeling, checking whether the integrated simulator reproduces multiple empirical response patterns in direction, shape, temporal dynamics, and trade-offs rather than matching single point estimates. Checkpointing supplies the counterfactual replay capability.
Core claim. MiniCorp couples two stateful systems—a self-advancing external world of customers, dynamic competitors, suppliers, and a marketplace platform, and a persistent internal firm of standing roles—so that a company's decisions propagate through explicit market mechanisms (search matching and ranking, sponsored-search auctions, click and conversion behavior, competitor responses) and produce lasting, sometimes delayed market effects. The paper's central claim is that this closed loop generates enterprise records connecting organizational decisions to their business context and downstream consequences, plus checkpointed counterfactual replays of the same state under different decisions, which static archives cannot provide. The demonstration is an e-commerce company, and the paper reports four end-to-end fidelity tests against empirical response patterns: a 10% price change yields arc elasticities of 2.9 and 3.2 near an empirical benchmark; a temporary four-week discount lifts units 71% with nearly half of the incremental sales occurring after price restoration; a two-week stockout leaves 59% of the cumulative unit shortfall occurring after restocking; and advertising returns $0.53 per ad dollar with 84% of incremental units coming from competitor displacement. These responses are presented as system-level evidence that the integrated world, not merely its individually grounded parts, reacts to business decisions the way real markets do.
The premise it rests on. The whole enterprise is useful only if the simulated market responds to business decisions the way real markets do, yet fidelity rests on four end-to-end tests run from a single checkpoint with no reported variance, and the authors state that robustness across independently initialized worlds remains to be evaluated.
If the paper is right (Pith's inferences, not the authors')
MiniCorp yields enterprise data with full provenance: every work item carries its owner and origin, every message its sender and recipient, and every decision the role that proposed it and the ruling it received, so an outcome is attributable to a role and a decision rather than to the firm as a whole.
Checkpointing supplies counterfactual pairs—the same state under alternative decisions—which enables interactive reinforcement learning that frozen historical archives cannot support.
Agents exhibit realistic organizational behavior: work arrives mostly from colleagues (58% carried over, 31% message- or interrupt-triggered), CEO rejection rates track financial exposure (29% for replenishment spend versus 13% for quality incidents), and the separation of capability from authority routes lateral coordination between Quality Engineering and Supply Chain.
Explicit long-term strategic guidance sustains advertising through weak early returns and carries the firm past its cold start: the guided firm retained $1,913 in gross profit after advertising spend, versus a $7 loss without guidance.
The firm can route the right repair to the right role under external shocks, fully answering 11 of 20 runs across ten business situations, with collaboration proving irreplaceable in four of them.
The four validated response patterns cover a small slice of the decision space; if the same interventions were re-run across many independently initialized worlds and the reported magnitudes (arc elasticity near 3, 59% stockout shortfall, 84% displacement share) collapsed or reversed direction, agents trained in MiniCorp would be learning simulator-specific regularities rather than transferable bu
Checkpointed replay combined with per-role provenance makes MiniCorp a natural testbed for credit assignment in multi-agent systems, since the records separate a wrong diagnosis from a correct diagnosis with an unproposed action, and either of those from a proposal the approving role declined.
The external-world construction recipe—economic mechanisms, empirical evidence, real marketplace data, and adaptive competitors—is the generalizable part of the contribution; the e-commerce firm is one instance of a family the framework could extend to other industries and organizational structures.
The title-rewriting episode that doubled a listing's weekly sales suggests agents can develop their own operating practices, but whether such practices transfer to real marketplaces is unverified and would require testing policies learned in simulation against real-world outcomes.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
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.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-06
The authors' main claim, as the classifier picked it out 1.00
“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.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.32 · human participants 0.05 · controlled comparison 0.10 · randomized assignment 0.02 · states quantitative result 0.04 · compares with prior results 0.34 · mentions code or data availability 0.06
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| The last mile toward enterprise AGI is a company that runs itself. | background 0.92 | 0.06 | 0.22 | asserted 1.00 |
| Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints. | gap 0.55 | 0.08 | 0.03 | asserted 1.00 |
| Historical archives are also frequently incomplete and record only what actually happened. | background 0.66 | 0.07 | 0.02 | asserted 0.90 |
| They cannot show the outcomes of alternative decisions. | gap 0.58 | 0.07 | 0.02 | asserted 1.00 |
| We introduce MiniCorp, an office simulator for studying how agents can collectively run a company while generating enterprise data at scale. | contribution 1.00 | 0.84 | 0.11 | asserted 0.99 |
| Using an e-commerce company as a demonstration, MiniCorp connects two interacting worlds. | method 0.80 | 0.65 | 0.03 | asserted 1.00 |
| The external world models customers, dynamic competitors, and market mechanisms. | method 0.96 | 0.48 | 0.02 | asserted 1.00 |
| The internal world consists of agents that observe events, discuss their options, and make strategic decisions. | method 0.97 | 0.31 | 0.03 | asserted 1.00 |
| These decisions have lasting effects on the market, and the resulting feedback informs the firm's later decisions. | method 0.99 | 0.31 | 0.05 | asserted 1.00 |
| As the firm and market interact, MiniCorp continuously records the agents' communications and decisions. | method 0.98 | 0.72 | 0.03 | asserted 1.00 |
| These records preserve the information available at the time and the business results that followed. | method 0.96 | 0.67 | 0.05 | asserted 1.00 |
| Checkpointing allows the same situation to be replayed under different decisions, providing comparisons unavailable in static archives. | method 0.91 | 0.61 | 0.08 | asserted 1.00 |
| We evaluate end-to-end fidelity against patterns reported in empirical studies of real markets. | method 0.95 | 0.89 | 0.04 | asserted 0.98 |
| These evaluations provide agents with realistic market feedback and reduce the risk that they learn to exploit flaws in the simulator. | method 0.71 | 0.71 | 0.06 | asserted 0.97 |
| Our experiments show agents coordinating across roles and adapting their decisions to market feedback. | result 1.00 | 0.94 | 0.10 | asserted 1.00 |
| With explicit long-term strategic guidance, they also sustain advertising exploration despite weak early returns. | result 1.00 | 0.91 | 0.07 | asserted 1.00 |
| 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. | contribution 0.76 | 0.63 | 0.88 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 0.86 · evidence from simulation or computation 0.77
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-06: 1 request, 11,674 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.

No comments yet.
Sign in to comment