spb/wp25_uqo

1# Literature matrix and novelty audit — UQO Working Paper No. 2523**Paper:** *When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory*4**Prepared:** 2026-10-27 by the literature-review / novelty-audit member of the team. **Status:** verified (every row below has a BibTeX entry in `paper/references.bib` whose bibliographic details were checked against Crossref, arXiv, NBER or the publisher page on 2026-10-27; see `docs/references_verification.md`). Citation keys in `code` are the BibTeX keys.56**Mechanism under audit (quoted from the research brief, not modified here):** delegation converts idiosyncratic decision noise into common, model-correlated noise. Humans act on private forecasts with independent errors that wash out by the law of large numbers; a user who delegates to an AI agent built on foundation model *m* gains forecast precision privately but inherits an error common to all users of *m*. Aggregates (capital accumulation, asset demand) then load on the shared model error with weight *D* (delegation share) and 1/*M* (number of independent models); prices inherit "model noise"; the private delegation decision ignores this pecuniary externality, so competitive adoption is excessive for finite *M*; model diversity or a Pigouvian delegation tax restores efficiency. A compute market links inference demand to task-based automation through a common compute rental price (rents to algorithmic-capital owners, wealth concentration); with collateral constraints, common model shocks are amplified through margin spirals.78**How to read the tables.** Columns: reference; venue; DOI/URL (verified); model class / equilibrium concept; role of AI or information; key mechanism; what WP25 adds or how it differs. The last column is written from the referee's chair: when a strand already contains a piece of the mechanism, the table says so.910---1112## 1. Arrow–Debreu / Radner equilibrium and constrained (in)efficiency1314| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |15|---|---|---|---|---|---|---|16| Arrow & Debreu 1954 `arrow1954existence` | Econometrica 22(3) | 10.2307/1907353 | Static competitive equilibrium; existence via fixed point | None (complete markets, full information) | Convexity + continuity give existence | WP25 needs existence with an endogenous delegation choice that changes the information structure; the Arrow–Debreu benchmark is the *D = 0* limit of the model. |17| Debreu 1959 `debreu1959theory` | Yale UP (Cowles Monograph 17) | yalebooks.yale.edu/book/9780300015591 | Axiomatic GE; uncertainty via state-contingent commodities | Information is a contingency structure, not a choice | Contingent-claims markets | WP25 makes the contingency structure (who shares which error) a market outcome. |18| Radner 1972 `radner1972existence` | Econometrica 40(2) | 10.2307/1909407 | Sequential equilibrium of plans, prices and price expectations | Agents hold heterogeneous information; expectations must be consistent | Existence under differential information and bounded trades | WP25's baseline is a Radner-type sequence economy; the novelty is not the concept but the endogenous correlation of forecast errors. |19| Radner 1979 `radner1979rational` | Econometrica 47(3) | 10.2307/1910413 | Fully revealing REE, generic existence | Prices aggregate dispersed information | Generic full revelation when states are finite | WP25 will need noise (model noise) to prevent full revelation, as in Grossman–Stiglitz; "model noise" is a *new source of non-revelation* only if it is not spanned by prices. |20| Lucas 1978 `lucas1978asset` | Econometrica 46(6) | 10.2307/1913837 | Recursive exchange economy, representative agent | Rational expectations | Euler-equation asset pricing | WP25's asset-pricing block nests Lucas when *D = 0*; with *D > 0* the SDF loads on a common model error. |21| Greenwald & Stiglitz 1986 `greenwald1986externalities` | QJE 101(2) | 10.2307/1891114 | GE with imperfect information / incomplete markets | Information frictions | Equilibria are generically constrained-inefficient because of pecuniary externalities | **The externality logic WP25 uses is this one.** WP25's contribution is only the *specific* externality (delegation changes the error covariance faced by others), not the existence of pecuniary externalities under incomplete markets. |22| Dávila & Korinek 2018 `davila2018pecuniary` | RES 85(1) | 10.1093/restud/rdx010 | GE with financial frictions; distributive vs collateral externalities | — | Decomposes pecuniary externalities into distributive and collateral terms; Pigouvian corrections | WP25's "delegation tax" must be derived with this decomposition; the collateral-amplification block is a direct application. |23| Lorenzoni 2008 `lorenzoni2008inefficient` | RES 75(3) | 10.1111/j.1467-937X.2008.00494.x | Three-period GE with limited commitment | — | Over-borrowing because agents ignore price effects of fire sales | Closest template for "excessive adoption": replace borrowing by delegation and fire-sale prices by model-noise-loaded prices. |2425## 2. Heterogeneous-agent incomplete-markets macro (Aiyagari / Huggett / Krusell–Smith) and HANK2627| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |28|---|---|---|---|---|---|---|29| Huggett 1993 `huggett1993riskfree` | JEDC 17(5–6) | 10.1016/0165-1889(93)90024-M | Stationary RCE, idiosyncratic endowment risk, borrowing limit | None | Precautionary saving lowers risk-free rate | Baseline for Model B; WP25 adds an *aggregate* shock that is created by delegation. |30| Aiyagari 1994 `aiyagari1994uninsured` | QJE 109(3) | 10.2307/2118417 | Stationary RCE with capital | None | Precautionary over-accumulation | Same. |31| Krusell & Smith 1998 `krusell1998income` | JPE 106(5) | 10.1086/250034 | RCE with aggregate + idiosyncratic risk; approximate aggregation | None | Distribution matters little for aggregates | WP25's model noise is an aggregate shock whose variance is endogenous (*D*, *M*); approximate aggregation may fail when delegation is wealth-dependent — this is a testable numerical claim. |32| Kaplan, Moll & Violante 2018 `kaplan2018monetary` | AER 108(3) | 10.1257/aer.20160042 | HANK, two assets | None | Indirect (GE) effects dominate direct effects of policy | WP25 should decompose delegation effects into direct (precision gain) and indirect (price/model-noise) channels in the same spirit. |33| Auclert, Bardóczy, Rognlie & Straub 2021 `auclert2021using` | Econometrica 89(5) | 10.3982/ECTA17434 | Sequence-space Jacobians for HA models | None | Fast linearised solution of HA models with aggregate shocks | Candidate solution method for Model B. |34| Achdou et al. 2022 `achdou2022income` | RES 89(1) | 10.1093/restud/rdab002 | Continuous-time HA (HJB + KF) | None | Mean-field-game formulation | Alternative solver; also the natural language for "a continuum of agents sharing a common noise" (common-noise MFG). |35| Hubmer, Krusell & Smith 2021 `hubmer2021sources` | NBER Macro Annual 35 | 10.1086/712332 | Quantitative HA with heterogeneous returns | None | Return heterogeneity drives top wealth shares | WP25's "ownership of algorithmic capital" is a return-heterogeneity channel; the H3 claim must be benchmarked against this. |3637## 3. Task-based automation and the macroeconomics of AI3839| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |40|---|---|---|---|---|---|---|41| Zeira 1998 `zeira1998workers` | QJE 113(4) | 10.1162/003355398555847 | Growth with task automation | AI as capital that replaces labour in tasks | Automation raises capital share | Origin of the task framework WP25 reuses; no information role. |42| Acemoglu & Restrepo 2018 `acemoglu2018race` | AER 108(6) | 10.1257/aer.20160696 | Task-based GE with endogenous automation and new tasks | AI as production technology (iv in the brief's typology) | Displacement vs reinstatement; wage and labour-share effects | WP25 keeps this production block but adds AI as *decision-maker*: the brief forbids a model whose only AI channel is production. |43| Acemoglu & Restrepo 2022 `acemoglu2022tasks` | Econometrica 90(5) | 10.3982/ECTA19815 | Task displacement and wage structure | Production | Task displacement explains most of the rise in US wage inequality | Empirical anchor for calibrating displacement. |44| Moll, Rachel & Restrepo 2022 `moll2022uneven` | Econometrica 90(6) | 10.3982/ECTA19417 | HA model with automation | Production | Automation raises returns to wealth → wealth inequality even with wage gains | **H3 is already partly here**: automation raising *r* concentrates wealth. WP25's increment is the *compute-rent* channel (scarce compute, concentrated ownership of algorithmic capital), which must be shown to be distinct from a generic rise in *r*. |45| Acemoglu 2025 `acemoglu2025simple` | Economic Policy 40(121) | 10.1093/epolic/eiae042 | Task-based growth accounting | Production + "easy-to-learn tasks" | Modest TFP gains from AI; negative "bad-task" effects possible | WP25's productivity calibration should be bounded by this. |46| Korinek & Suh 2024 `korinek2024scenarios` | NBER w32255 | 10.3386/w32255 | Task-complexity distribution; AGI scenarios | Production | Wages collapse when compute overtakes the task-complexity frontier | Compute scarcity as the binding constraint — overlaps with WP25's H6; WP25 adds the *inference-vs-automation competition for compute* through one rental price. |47| Restrepo 2025 `restrepo2025we` | NBER w34423 (chapter in *Economics of Transformative AI*) | 10.3386/w34423 | Growth with bottleneck vs supplementary work | Production (compute as input) | With abundant compute all bottleneck work is automated; wages → cost of replicating human work; labour share → 0 | WP25's compute market is a *market-clearing* version of Restrepo's compute constraint, with inference demand from delegation competing with automation demand. |48| Aghion, Jones & Jones 2019 `aghion2019artificial` | in *Economics of AI: An Agenda* (U. Chicago Press) | 10.7208/chicago/9780226613475.003.0009 | Growth with AI in production and in idea production; Baumol bottlenecks | Production / R&D | Bottleneck tasks limit growth; singularity conditions | Growth-side benchmark; WP25 is not a growth paper. |49| Jones 2024 `jones2024dilemma` | AER: Insights 6(4) | 10.1257/aeri.20230570 | Growth vs existential risk | — | Welfare trade-off of AI adoption under catastrophic risk | Different risk (existential); WP25's risk is correlated error. |50| Jones 2025 `jones2025how` | NBER w33602 | 10.3386/w33602 | Same, quantitative | — | Willingness to pay to reduce AI risk | Same. |51| Nordhaus 2021 `nordhaus2021are` | AEJ: Macro 13(1) | 10.1257/mac.20170105 | Growth with IT capital | Production | Tests for singularity | Background. |52| Ide & Talamàs 2025 `ide2025artificial` | JPE 133(12) | 10.1086/737233 | Knowledge-hierarchy GE | AI as problem-solver in organisations | Autonomous vs non-autonomous AI reshape hierarchies and wages | Closest "AI as decision-maker in GE" paper on the labour side; **no correlated-error or model-count channel**. |53| Agrawal, Gans & Goldfarb 2019 `agrawal2019artificial` | JEP 33(2) | 10.1257/jep.33.2.31 | Task = prediction + judgement | AI as *prediction* technology | Cheaper prediction substitutes for human prediction, complements judgement | WP25's "forecast precision gain from delegation" is the formalisation of this; what AGG do not model is that the prediction is *shared*. |54| Athey, Bryan & Gans 2020 `athey2020allocation` | AEA P&P 110 | 10.1257/pandp.20201034 | Principal–agent with AI vs human decision authority | AI as delegate | Allocation of decision rights between human and AI under misaligned objectives | Partial-equilibrium delegation threshold; **WP25's H1 threshold in GE extends this**, with the externality absent in AB&G. |55| Hadfield & Koh 2025 `hadfield2025economy` | arXiv 2509.01063 (chapter) | arxiv.org/abs/2509.01063 | Verbal survey; no formal propositions | AI agents as consumers/producers | States that GE distortions are precluded if markets are large **and AI mistakes are zero-mean and independent**; notes AI errors may be more correlated because the same agent is "copied" within and between firms, "inducing correlated mistakes that do not wash out in the aggregate" (their §2.1, §3.4) | **This is the verbal statement of WP25's core premise.** WP25 must cite it and position itself as the formalisation: endogenous *D*, the 1/*M* loading, the externality, the price consequences and the policy result are all absent in Hadfield–Koh. |56| Rothschild, Mobius, Hofman & Dillon 2025 `rothschild2025agentic` | arXiv 2505.15799 | arxiv.org/abs/2505.15799 | Verbal | AI agents as delegates in consumer–business matching | Delegation lowers communication frictions | No correlated-error channel. |57| Catalini, Hui & Wu 2026 `catalini2026some` | arXiv 2602.20946 | arxiv.org/abs/2602.20946 | Simple analytical AGI economics | Production | Allocation and pricing of AGI capabilities | Background for the compute-price block. |58| Eloundou et al. 2024 `eloundou2024gpts` | Science 384(6702) | 10.1126/science.adj0998 | Task exposure measurement | LLM exposure | ~80 % of workers have ≥10 % of tasks exposed | Calibration of the delegable-task share. |59| Brynjolfsson, Li & Raymond 2025 `brynjolfsson2025generative` | QJE 140(2) | 10.1093/qje/qjae044 | Field experiment | Generative AI as assistant | +14 % productivity, largest for novices | Calibrates the private precision gain *and* its heterogeneity (novices gain more → delegation is wealth/skill-dependent). |60| Eisfeldt, Schubert & Zhang 2023 `eisfeldt2023generative` | NBER w31222 | 10.3386/w31222 | Event study | GenAI exposure of firms | Exposed firms gained value after ChatGPT | Evidence that AI exposure is priced; WP25 predicts a *common* factor in exposed-firm returns. |61| Bick, Blandin & Deming 2024 `bick2024rapid` | NBER w32966 | 10.3386/w32966 | Survey | GenAI adoption | 39 % of US adults used genAI by Aug 2024 | Calibrates *D*. |6263## 4. Endogenous growth6465| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |66|---|---|---|---|---|---|---|67| Romer 1990 `romer1990endogenous` | JPE 98(5, Pt 2) | 10.1086/261725 | Monopolistic competition, non-rival ideas | — | Ideas non-rival → increasing returns | Foundation-model weights are non-rival ideas with *excludable* access (rents) — WP25's algorithmic capital. |68| Aghion & Howitt 1992 `aghion1992model` | Econometrica 60(2) | 10.2307/2951599 | Schumpeterian growth | — | Creative destruction | Background. |69| Jones 1995 `jones1995rd` | JPE 103(4) | 10.1086/262002 | Semi-endogenous growth | — | Scale effects removed | Background. |70| Jones & Tonetti 2020 `jones2020nonrivalry` | AER 110(9) | 10.1257/aer.20191330 | GE with data as non-rival input | Data | Who owns data determines sharing and growth; privacy externality | Ownership of a non-rival information input matters for allocation — WP25's model-ownership block should cite this as the data analogue. |7172## 5. Principal–agent theory and delegation7374| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |75|---|---|---|---|---|---|---|76| Holmström 1979 `holmstrom1979moral` | Bell J. Econ. 10(1) | 10.2307/3003320 | Moral hazard, sufficient statistic | — | Informativeness principle | WP25's "AI agents are constrained policies" is a contract on observables; informativeness principle decides what the mandate conditions on. |77| Holmström & Milgrom 1987 `holmstrom1987aggregation` | Econometrica 55(2) | 10.2307/1913238 | Dynamic moral hazard, linear contracts | — | Linear contracts optimal under Brownian output | Tractable mandate form (linear risk limits). |78| Holmström & Milgrom 1991 `holmstrom1991multitask` | JLEO 7 (special) | 10.1093/jleo/7.special_issue.24 | Multitask agency | — | Incentives distort effort across tasks | Multi-task AI mandates (trading + risk) distort. |79| Aghion & Tirole 1997 `aghion1997formal` | JPE 105(1) | 10.1086/262063 | Formal vs real authority | Information acquisition | Delegation raises agent initiative, loses principal control | WP25's H1 is an Aghion–Tirole delegation decision with a GE externality attached. |80| Dessein 2002 `dessein2002authority` | RES 69(4) | 10.1111/1467-937X.00227 | Delegation vs cheap-talk communication | — | Delegate when bias is small relative to information | Delegation threshold template. |81| Alonso & Matouschek 2008 `alonso2008optimal` | RES 75(1) | 10.1111/j.1467-937X.2007.00471.x | Optimal delegation sets | — | Interval delegation | Mandate = delegation set; WP25's "principal-approved constraints". |82| D'Acunto, Prabhala & Rossi 2019 `dacunto2019promises` | RFS 32(5) | 10.1093/rfs/hhz014 | Empirical, robo-advisor adoption | AI (robo) as delegate | Adopters diversify more; under-diversified investors gain | Evidence on who delegates; WP25 predicts robo users' trades become *correlated* — testable. |83| Rossi & Utkus 2024 `rossi2024diversification` | JFE 157 | 10.1016/j.jfineco.2024.103869 | Empirical | Robo-advice | Welfare gains from robo-advice | Same. |8485## 6. Financial intermediation, collateral and leverage8687| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |88|---|---|---|---|---|---|---|89| Kiyotaki & Moore 1997 `kiyotaki1997credit` | JPE 105(2) | 10.1086/262072 | Collateral-constrained GE | — | Asset price ↔ borrowing feedback amplifies shocks | WP25's H4 is Kiyotaki–Moore with the shock being model noise; the amplification is *not* new, the shock's origin is. |90| Bernanke, Gertler & Gilchrist 1999 `bernanke1999financial` | Handbook of Macro 1C | 10.1016/S1574-0048(99)10034-X | DSGE with costly state verification | — | Financial accelerator | Same. |91| Geanakoplos 2010 `geanakoplos2010leverage` | NBER Macro Annual 24 | 10.1086/648285 | GE with endogenous leverage, heterogeneous optimism | Heterogeneous beliefs | Leverage cycle | WP25's margin spiral under common model shocks; note Geanakoplos needs *heterogeneity*, WP25's model noise *reduces* heterogeneity — must reconcile. |92| Brunnermeier & Pedersen 2009 `brunnermeier2009market` | RFS 22(6) | 10.1093/rfs/hhn098 | Margin-constrained speculators | — | Funding–market liquidity spiral | Direct template for the amplification block. |93| Gertler & Kiyotaki 2010 `gertler2010financial` | Handbook of Monetary Econ. 3 | 10.1016/B978-0-444-53238-1.00011-9 | DSGE with bank balance sheets | — | Bank net worth and credit policy | Bank block of Model C. |94| Brunnermeier & Sannikov 2014 `brunnermeier2014macroeconomic` | AER 104(2) | 10.1257/aer.104.2.379 | Continuous-time macro-finance | — | Volatility paradox, non-linear amplification | Model C's non-linear dynamics. |95| Simsek 2013 `simsek2013belief` | Econometrica 81(1) | 10.3982/ECTA9956 | Collateral with belief disagreement | Beliefs | Collateral constraints discipline optimism asymmetrically | Belief homogenisation (WP25) shuts this channel — a competing effect: fewer disagreements may *reduce* leverage. |9697## 7. Rational-expectations information aggregation and market microstructure9899| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |100|---|---|---|---|---|---|---|101| Grossman & Stiglitz 1980 `grossman1980impossibility` | AER 70(3) | RePEc:aea:aecrev:v:70:y:1980:i:3:p:393-408 | Noisy REE with costly information | Costly private signals | Prices cannot be fully revealing if information is costly | WP25's delegation fee is an information cost; **"model noise" must be shown to be a non-price-revealed noise**, else prices reveal it and the externality vanishes. |102| Hellwig 1980 `hellwig1980aggregation` | JET 22(3) | 10.1016/0022-0531(80)90056-3 | Large-economy noisy REE | Dispersed signals | Aggregation of independent signals in prices | The *independent-error* benchmark WP25 departs from. |103| Diamond & Verrecchia 1981 `diamond1981information` | JFE 9(3) | 10.1016/0304-405X(81)90026-X | Noisy REE with idiosyncratic endowment noise | — | Same | Same. |104| Kyle 1985 `kyle1985continuous` | Econometrica 53(6) | 10.2307/1913210 | Strategic insider, market maker | Private information | Price impact λ | Market-design block (continuous double auction) and the Kyle setting used by Meng & Chen 2026 and Park 2026 (see §10). |105| Vives 2008 `vives2008information` | Princeton UP | 10.1515/9781400829507 | Monograph | — | Information aggregation and learning in markets | Reference text. |106| Vives 2014 `vives2014possibility` | JEEA 12(5) | 10.1111/jeea.12107 | REE with endogenous information | — | Conditions under which markets can be informationally efficient | Background. |107| Morris & Shin 2002 `morris2002social` | AER 92(5) | 10.1257/000282802762024610 | Beauty-contest game with public + private signals | Public (common) signal | Agents over-weight the common signal; more precise public information can lower welfare | **Common-signal over-reaction is old.** WP25's increment: the "public signal" is a *purchased* model output with endogenous market share, and the loss operates through prices rather than coordination payoffs. |108| Angeletos & Pavan 2007 `angeletos2007efficient` | Econometrica 75(4) | 10.1111/j.1468-0262.2007.00783.x | General class with common/private information | Information structure | Efficient use of information; welfare value of common information depends on strategic complementarity | Provides the welfare machinery; WP25's constrained optimum must be stated in these terms. |109110## 8. Dispersed information, sentiments, information choice and data111112| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |113|---|---|---|---|---|---|---|114| Lorenzoni 2009 `lorenzoni2009theory` | AER 99(5) | 10.1257/aer.99.5.2050 | DSGE with dispersed information and a noisy public signal | Common noisy signal about productivity | Noise shocks to a common signal generate demand-driven fluctuations | **"Common signal noise causes aggregate fluctuations" is already here.** WP25's model error is Lorenzoni's noise shock with an endogenous variance (*D*, *M*) and an endogenous private decision to listen to it. |115| Angeletos & La'O 2013 `angeletos2013sentiments` | Econometrica 81(2) | 10.3982/ECTA10008 | Unique-equilibrium business cycle with correlated beliefs | Communication frictions | "Sentiment" shocks = extrinsic correlated beliefs drive fluctuations without fundamentals | **The closest macro antecedent**: correlated belief errors cause fluctuations. WP25's claim must be that the correlation is *produced by a market for delegation* and is *inefficiently high*; A–L's correlation is exogenous. |116| Benhabib, Wang & Wen 2015 `benhabib2015sentiments` | Econometrica 83(2) | 10.3982/ECTA11085 | Sentiment-driven equilibria via signal extraction | — | Self-fulfilling sentiment equilibria | Alternative route to sentiment fluctuations; WP25 should check whether model noise creates multiplicity. |117| David, Hopenhayn & Venkateswaran 2016 `david2016information` | QJE 131(2) | 10.1093/qje/qjw006 | Firm investment under imperfect information; misallocation | Firm forecasts | Forecast errors → misallocation → TFP loss | WP25's firm block: delegation raises each firm's forecast precision (less misallocation) but correlates errors (more aggregate volatility) — the trade-off is the new object. |118| Sims 2003 `sims2003implications` | JME 50(3) | 10.1016/S0304-3932(03)00029-1 | Rational inattention | Capacity-constrained information | Attention limits create sluggish, noisy responses | Delegation as outsourcing attention; could microfound the precision gain. |119| Veldkamp 2011 `veldkamp2011information` | Princeton UP | 10.1515/9781400840496 | Monograph on information choice | — | Information choice in macro and finance | Reference text. |120| Veldkamp 2006 `veldkamp2006media` | AER 96(3) | 10.1257/aer.96.3.577 | Markets for information with fixed costs | Information as a good with increasing returns | Many agents buy the *same* information → herding and media frenzies | **Endogenous purchase of common information is here.** WP25 differs in that the shared object is a *decision rule* (agent), not a news item, and in the GE-price externality. |121| Hellwig & Veldkamp 2009 `hellwig2009knowing` | RES 76(1) | 10.1111/j.1467-937X.2008.00515.x | Information acquisition in coordination games | — | Complementarity in actions → complementarity in information choice → agents choose the same signals | **Coordination-driven homogenisation of information sources is here.** WP25 needs a reason for correlated adoption that is not strategic complementarity (e.g., scale economies in model training) and must say so. |122| Colombo, Femminis & Pavan 2014 `colombo2014information` | RES 81(4) | 10.1093/restud/rdu015 | Endogenous private-information acquisition, welfare | — | Inefficiency in both use and acquisition of information | Welfare benchmark for an information-acquisition externality — WP25's delegation externality is of this family. |123| Vives 2017 `vives2017endogenous` | RES 84(2) | 10.1093/restud/rdw062 | Market games with endogenous public information | Public information from prices | Welfare effects of public information in market games | Same family. |124| Farboodi & Veldkamp 2020 `farboodi2020long` | AER 110(8) | 10.1257/aer.20171349 | Growth of data technology in a noisy-REE market | Data-processing technology as a choice | Shift from fundamental to order-flow information as data tech improves | **Information-technology choice is here.** WP25 differs by making the technology *shared* across users and counting independent technologies (*M*). |125| Farboodi, Matray, Veldkamp & Venkateswaran 2022 `farboodi2022where` | RFS 35(7) | 10.1093/rfs/hhab110 | Same framework, cross-section | Data | Data accrues to large, growth firms | Empirical background. |126| Farboodi & Veldkamp 2023 `farboodi2023data` | Annual Rev. Econ. 15 | 10.1146/annurev-economics-082322-023244 | Survey | Data economics | Data as by-product, non-rival, feedback | Survey to cite. |127| Dugast & Foucault 2018 `dugast2018data` | JFE 130(2) | 10.1016/j.jfineco.2018.07.004 | Noisy REE with raw vs processed data | Data abundance | Cheap raw data can *reduce* long-run price informativeness | A competing explanation for lower informativeness under AI that does not need correlated errors. |128129## 9. Heterogeneous beliefs and speculation130131| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |132|---|---|---|---|---|---|---|133| Miller 1977 `miller1977risk` | J. Finance 32(4) | 10.1111/j.1540-6261.1977.tb03317.x | Short-sale constraints + divergence of opinion | — | Prices reflect optimists → overpricing | Belief homogenisation removes the Miller premium — competing prediction (lower prices/less overvaluation) to WP25's "model noise raises volatility". |134| Harrison & Kreps 1978 `harrison1978speculative` | QJE 92(2) | 10.2307/1884166 | Dynamic heterogeneous beliefs | — | Resale option → speculative premium | Same. |135| Scheinkman & Xiong 2003 `scheinkman2003overconfidence` | JPE 111(6) | 10.1086/378531 | Overconfidence-driven disagreement | — | Bubbles and volume from disagreement | Same: WP25 must confront the prediction that fewer independent models reduce disagreement-driven volume and bubbles. |136| Hong & Stein 2007 `hong2007disagreement` | JEP 21(2) | 10.1257/jep.21.2.109 | Survey | — | Disagreement models | Survey. |137138## 10. Algorithmic / AI trading, herding, homogenisation and financial stability139140| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |141|---|---|---|---|---|---|---|142| Banerjee 1992 `banerjee1992simple` | QJE 107(3) | 10.2307/2118364 | Sequential social learning | — | Informational cascades | Herding by *observation*; WP25's herding is by *shared technology*, no observation needed. |143| Chaboud et al. 2014 `chaboud2014rise` | J. Finance 69(5) | 10.1111/jofi.12186 | Empirical, FX algorithmic trading | Algorithms | Algo strategies are *more correlated* than human strategies, yet improve efficiency | **Empirical precedent for correlated algorithmic trades** (pre-LLM). WP25 should use this as the pre-foundation-model baseline. |144| Kirilenko, Kyle, Samadi & Tuzun 2017 `kirilenko2017flash` | J. Finance 72(3) | 10.1111/jofi.12498 | Empirical, Flash Crash | HFT | Synchronised withdrawal of liquidity | Stress-time evidence for H2. |145| Calvano et al. 2020 `calvano2020artificial` | AER 110(10) | 10.1257/aer.20190623 | Q-learning pricing agents in oligopoly | RL agents | Autonomous algorithms learn to collude | Alternative mechanism for correlated AI behaviour (learning), not shared errors. |146| Dou, Goldstein & Ji 2025 `dou2025ai` | NBER w34054 | 10.3386/w34054 | Informed-speculator model + RL simulations | RL trading agents | Collusion via price-trigger strategies and via **homogenised learning biases ("artificial stupidity")**; lower informativeness and liquidity | **Homogenised AI biases reducing price efficiency is here**, in a Kyle-type setting with learning agents. WP25 differs: errors come from a *shared forecasting model*, adoption is a *household/firm* decision, and the consequences are macro (capital, wages, *r*). |147| Gabaix & Koijen 2021 `gabaix2021search` | NBER w28967 | 10.3386/w28967 | Inelastic-markets GE | — | Flows move prices with multiplier ≈ 5 | Transmission from correlated flows to prices; WP25 needs an inelastic market for model noise to matter for prices. |148| Gârleanu & Pedersen 2018 `garleanu2018efficiently` | J. Finance 73(4) | 10.1111/jofi.12696 | Asset-management equilibrium with costly search | Delegation to asset managers | Delegation and market efficiency jointly determined | Delegation to *human* managers already affects efficiency; WP25's distinct claim is the error-correlation structure of *AI* managers. |149| Khandani & Lo 2011 `khandani2011what` | J. Fin. Markets 14(1) | 10.1016/j.finmar.2010.07.005 | Empirical, Aug-2007 quant crisis | Quant strategies | Crowded, similar quant strategies unwound together | **Empirical precedent for "same model → same trades → correlated losses"** (quant funds, 2007). |150| Stein 2009 `stein2009presidential` | J. Finance 64(4) | 10.1111/j.1540-6261.2009.01472.x | Arbitrage with crowding and leverage | Sophisticated investors | Crowded-trade and leverage externalities: each arbitrageur ignores the effect of his trade on others' unwinding | **The crowding externality is here, verbally and in a simple model.** WP25's delegation externality is a crowding externality in forecast space; must cite and differentiate (endogenous *M*, macro aggregates). |151| Wagner 2011 `wagner2011systemic` | J. Finance 66(4) | 10.1111/j.1540-6261.2011.01666.x | Portfolio choice with liquidation costs | — | Diversification by all makes portfolios *similar* → joint liquidation risk; diversity–diversification trade-off | **Private-gain/common-exposure trade-off with an inefficiency is here.** WP25's "precision gain vs error correlation" is isomorphic; the increment is the economic object (forecast models) and the compute/ownership links. |152| Kleinberg & Raghavan 2021 `kleinberg2021algorithmic` | PNAS 118(22) | 10.1073/pnas.2018340118 | Selection model with shared algorithm | Shared algorithm | Monoculture: even a more accurate shared algorithm can lower *aggregate* outcome quality (lost independence) | **"Shared accurate algorithm, lower aggregate quality" is here**, in a matching/selection context without prices. WP25 adds prices, dynamics and endogenous adoption. |153| Bommasani et al. 2022 `bommasani2022picking` | NeurIPS 2022 / arXiv 2211.13972 | arxiv.org/abs/2211.13972 | Empirical ML | Shared foundation models | Outcome homogenisation: same individuals fail across systems sharing a base model | Micro evidence that shared base models correlate errors across deployments. |154| Jo, Creel, Wilson & Raghavan 2025 `jo2025homogeneous` | NeurIPS 2025 / arXiv 2503.15634 | arxiv.org/abs/2503.15634 | Pricing game with correlated predictors | Shared pre-trained models | Correlated predictions soften competition; firms trade accuracy for coordination | Correlated AI predictions have equilibrium effects in product markets — adjacent, not macro. |155| Danielsson, Macrae & Uthemann 2022 `danielsson2022artificial` | J. Banking & Finance 140 | 10.1016/j.jbankfin.2021.106290 | Conceptual | AI in risk management and regulation | AI monoculture, procyclicality, opacity → systemic risk | Verbal statement of the monoculture concern. |156| Danielsson & Uthemann 2024 `danielsson2024artificial` | SSRN 4903998 | ssrn.com/abstract=4903998 | Conceptual | AI and crises | AI may amplify crises through similar reactions | Same. |157| Gensler & Bailey 2020 `gensler2020deep` | SSRN 3723132 | ssrn.com/abstract=3723132 | Policy essay | Deep learning in finance | Data/model concentration and "herding" from shared models threaten stability | Earliest explicit policy statement of model-monoculture risk; verbal. |158| FSB 2024 `fsb2024financial` | FSB report (Nov 2024) | fsb.org/2024/11/… | Policy | AI in finance | Third-party concentration, herding, model risk | Policy framing. |159| IMF 2024 `imf2024advances` | GFSR Oct 2024, Ch. 3 | imf.org/…/global-financial-stability-report-october-2024 | Policy + survey | AI in capital markets | Faster markets, higher volatility under stress, reliance on few AI providers | Policy framing; WP25's *M* is "number of AI providers". |160| BIS 2024 `bis2024artificial` | BIS Annual Economic Report 2024, Ch. III | bis.org/publ/arpdf/ar2024e3.htm | Policy | AI and the economy | Herding, runs and fire sales may be amplified | Policy framing. |161| Meng & Chen 2026 `meng2026artificial` | arXiv 2604.03272 | arxiv.org/abs/2604.03272 | Extended REE (Kyle-type) with **endogenous adoption**; supermodular adoption game | AI traders with a common signal component; performative feedback | Systemic-risk coupling depends on adoption share φ, signal correlation, feedback and endogenous price impact; multiplier convex in φ; tipping to algorithmic monoculture; 13F evidence (tail-loss amplification 18–54 %) | **Closest formal competitor in the market block**: endogenous adoption + correlated signals + superlinear systemic risk. Not in M&C: households/firms and macro prices (wages, *r*, capital), the 1/*M* model-count structure, the welfare/Pigouvian result, compute market. Abstract-level reading only (see verification file). |162| Qiu & Han 2026 `qiu2026representation` | arXiv 2604.22818 | arxiv.org/abs/2604.22818 | Calibrated multi-agent market model | AI agents with similar representations | Representation homogeneity compresses forecast disagreement under stress; hidden leverage; policy = preserve diversity | Simulation-based version of H2/H5 ("model diversity"); no adoption choice, no macro. |163| Spiro 2026 `spiro2026oracle` | arXiv 2605.00844 | arxiv.org/abs/2605.00844 | Empirical forecasting study | GPT-4o/Claude/Gemini forecasts | LLM errors strongly correlated *across* vendors (r ≈ 0.77); crowd independence "built but not yet activated" | **Challenges WP25's 1/*M* assumption**: independent vendors may not mean independent errors. WP25 must treat *M* as *effective* number of independent models, calibrated, not nominal. |164| Park 2026 `park2026correlated` | SSRN 7316679 | ssrn.com/abstract=7316679 | Noisy REE with correlated AI signals | AI homogenisation | Error correlation lowers price informativeness and price impact (per indexed abstract; full text not accessible on 2026-10-27) | Overlaps with WP25's asset-market proposition on correlated errors; WP25's increment is the delegation choice and the macro loop. |165| He 2026 `he2026when` | SSRN 6350140 | ssrn.com/abstract=6350140 | Empirical, ChatGPT outages as shocks | GenAI for retail investors | ChatGPT acts as a "belief homogeniser": synchronised retail trades, overreaction then reversal, worse liquidity | **Direct empirical evidence for the belief-homogenisation premise** — WP25 should cite as validation target. |166| Hansen & Lee 2025 `hansen2025financial` | Fed FEDS 2025-090 | 10.17016/FEDS.2025.090 | Lab-experiment replication with LLM agents | LLM agents in herding games | LLM agents herd *less* than humans, relying on private information; can be induced to herd when told to maximise profit | **Counter-evidence**: LLM agents do not herd by observation. Not inconsistent with WP25 (whose channel is shared *technology*, not observation) but must be discussed. |167| Lopez-Lira & Tang 2023 `lopezlira2023can` | SSRN 4412788 / arXiv 2304.07619 | ssrn.com/abstract=4412788 | Empirical | ChatGPT return forecasts | LLM sentiment predicts returns | Evidence that a single public model produces a tradable (hence shared) signal. |168| Doshi & Hauser 2024 `doshi2024generative` | Science Advances 10(28) | 10.1126/sciadv.adn5290 | Experiment | GenAI assistance | Individual quality up, collective diversity down | **Exact micro analogue of "private gain, collective homogenisation"** outside finance. |169170## 11. Agent-based computational economics171172| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |173|---|---|---|---|---|---|---|174| Tesfatsion 2006 `tesfatsion2006agent` | Handbook of Comp. Econ. 2 | 10.1016/S1574-0021(05)02016-2 | ACE methodology | Boundedly rational agents | Constructive equilibrium | WP25's Model C must state when a simulated outcome is *not* an equilibrium — Tesfatsion's constructive stance is the reference. |175| LeBaron 2006 `lebaron2006agent` | Handbook of Comp. Econ. 2 | 10.1016/S1574-0021(05)02024-1 | Agent-based finance | Learning traders | Stylised facts from heterogeneous learning | Benchmark for Model C stylised facts. |176| Hommes 2006 `hommes2006heterogeneous` | Handbook of Comp. Econ. 2 | 10.1016/S1574-0021(05)02023-X | Heterogeneous-agent models | Belief switching | Endogenous belief heterogeneity | Belief heterogeneity is *endogenous* here — WP25's delegation reduces it. |177| Brock & Hommes 1998 `brock1998heterogeneous` | JEDC 22(8–9) | 10.1016/S0165-1889(98)00011-6 | Adaptive belief systems | Fitness-based predictor choice | Routes to chaos as agents switch predictors | **Endogenous choice among forecasting rules is here** (predictor choice by past fitness). WP25's delegation is predictor choice with a *shared* predictor and a GE externality. |178| Farmer & Foley 2009 `farmer2009economy` | Nature 460 | 10.1038/460685a | Manifesto | — | ABM for macro | Motivation. |179| Dawid & Delli Gatti 2018 `dawid2018agent` | Handbook of Comp. Econ. 4 | 10.1016/bs.hescom.2018.02.006 | Agent-based macro survey | — | Macro ABMs | Survey. |180| Axtell & Farmer 2025 `axtell2025agent` | JEL 63(1) | 10.1257/jel.20221319 | Survey | — | State of ABM | Survey. |181| Gode & Sunder 1993 `gode1993allocative` | JPE 101(1) | 10.1086/261868 | Zero-intelligence traders in a CDA | No intelligence | Market institution, not rationality, yields efficiency | Control for Model C: ZI traders vs LLM traders isolates what the *agent* contributes. |182183## 12. Generative agents and LLM economic simulations184185| Reference | Venue | DOI / URL | Model class / equilibrium | Role of AI / information | Key mechanism | What WP25 adds / differs |186|---|---|---|---|---|---|---|187| Park et al. 2023 `park2023generative` | UIST '23 | 10.1145/3586183.3606763 | Generative agents with memory/reflection | LLM agents | Believable social simulation | Architecture template; no economics. |188| Horton 2023 `horton2023large` | NBER w31122 | 10.3386/w31122 | LLMs as *homo silicus* | LLM subjects | Replicates classic experiments | Methodology; warns LLMs are not humans. |189| Chen, Liu, Shan & Zhong 2023 `chen2023emergence` | PNAS 120(51) | 10.1073/pnas.2316205120 | Revealed-preference tests | GPT choices | GPT choices are GARP-consistent | Supports treating LLM agents as (constrained) optimisers. |190| Mei, Xie, Yuan & Jackson 2024 `mei2024turing` | PNAS 121(9) | 10.1073/pnas.2313925121 | Behavioural games | ChatGPT | Behaviour within human distribution, more cooperative | Same. |191| Ross, Kim & Lo 2024 `ross2024llm` | arXiv 2408.02784 | arxiv.org/abs/2408.02784 | Utility-theory elicitation | LLMs | LLMs have systematic, model-specific biases | **Model-specific bias = the "model error" WP25 posits**; supports *M*-indexed errors. |192| Bybee 2023 `bybee2023surveying` | arXiv 2305.02823 | arxiv.org/abs/2305.02823 | LLM-generated expectations | LLM beliefs | LLM expectations track survey expectations | LLM beliefs as a common belief-generating process. |193| Li et al. 2024 `li2024econagent` | ACL 2024 | 10.18653/v1/2024.acl-long.829 | Macro ABM with LLM households | LLM agents | Reproduces Phillips curve, Okun's law | Macro LLM simulation precedent; agents are homogeneous in model — WP25's *M*-experiment varies this. |194| Zhao et al. 2023 `zhao2023competeai` | ICML 2024 / arXiv 2310.17512 | arxiv.org/abs/2310.17512 | LLM firms competing | LLM agents | Competition dynamics emerge | Precedent. |195| Lopez-Lira 2025 `lopezlira2025can` | SSRN 5217340 / arXiv 2504.10789 | ssrn.com/abstract=5217340 | LLM traders in an order-book market | LLM agents | Strategy adherence; price discovery, bubbles; warns that similar architectures may behave in correlated ways | **The agentic-market experiment with a correlated-behaviour warning is here** (qualitative). WP25's experiments must *measure* correlation vs *M* and link to a theory. |196| Ross et al. 2026 `ross2026why` | arXiv 2609.04373 | arxiv.org/abs/2609.04373 | LLM agents + noise traders + market maker | Single-LLM-family populations | Higher LLM participation improves price discovery; "better models, riskier systems" framing | Agentic-market experiment on homogeneity; results are mixed — WP25 must engage. |197| Xia et al. 2026 `xia2026agentic` | arXiv 2605.19337 | arxiv.org/abs/2605.19337 | Survey | LLM trading agents | Flags strategy homogenisation from shared foundation models as open problem | Survey confirming the gap is recognised but unmodelled. |198| Fish, Gonczarowski & Shorrer 2024 `fish2024algorithmic` | arXiv 2404.00806 | arxiv.org/abs/2404.00806 | LLM pricing agents | LLM agents | LLMs collude in pricing | Alternative correlated-behaviour channel (collusion), not shared errors. |199| Immorlica, Lucier & Slivkins 2024 `immorlica2024generative` | ACM SIGecom Exchanges 22(1) | 10.1145/3699824.3699832 | Research agenda | GenAI as economic agents | Agents as delegates in mechanisms; alignment and competition | Agenda-level; no GE. |200| Hammond et al. 2025 `hammond2025multiagent` | arXiv 2502.14143 | arxiv.org/abs/2502.14143 | Risk taxonomy | Multi-agent AI | Correlated failures from shared models as a multi-agent risk | Taxonomy; names the risk, no economics. |201202---203204## Novelty audit205206### A. The twelve closest papers and what they already show2072081. **Angeletos & La'O 2013 (Econometrica)** — correlated belief errors ("sentiments") generate aggregate fluctuations in a unique-equilibrium RBC economy. *Already shows:* common noise in beliefs is an aggregate shock with real effects. *Gap:* the correlation is exogenous; no agent chooses to acquire a shared belief; no adoption externality; no policy on the number of belief sources.2092. **Lorenzoni 2009 (AER)** — noise in a common public signal produces demand-driven cycles. *Already shows:* "shared noisy signal → aggregate fluctuations". *Gap:* signal is free and public; no delegation decision; variance is a parameter, not an equilibrium object.2103. **Morris & Shin 2002 (AER); Angeletos & Pavan 2007 (Econometrica)** — agents over-weight common information; welfare can fall with more precise public information; full characterisation of efficient information use. *Already shows:* the social value of common information can be negative. *Gap:* WP25's loss runs through market prices and capital accumulation in a dynamic GE, not through a beauty-contest payoff; the information is purchased, rival in cost, and non-rival in content.2114. **Hellwig & Veldkamp 2009 (RES); Veldkamp 2006 (AER)** — endogenous information choice with complementarity leads agents to buy the same information; markets for information with fixed costs create frenzies. *Already shows:* homogenisation of information sources is an equilibrium outcome. *Gap:* the driver is strategic complementarity or news fixed costs; in WP25 it should be the scale economies of foundation models plus a private precision gain, with no need for complementarity — this difference must be made explicit, otherwise the result is a relabelling.2125. **Farboodi & Veldkamp 2020 (AER)** — choice of information *technology* in a noisy-REE market; the technology changes what prices reveal. *Already shows:* information-technology choice is an equilibrium object with market-efficiency effects. *Gap:* technologies are private; no shared-technology error; no count *M*.2136. **Stein 2009 (J. Finance); Wagner 2011 (J. Finance)** — crowding and leverage externalities; diversity–diversification trade-off (private diversification makes portfolios similar and raises joint-liquidation risk; the market under-provides diversity). *Already shows:* "private benefit, common exposure, inefficient similarity" with a diversity-restoring policy. *Gap:* the object is portfolios/strategies, not forecasting models; no macro aggregates; no endogenous *M*.2147. **Kleinberg & Raghavan 2021 (PNAS); Bommasani et al. 2022** — algorithmic monoculture: a shared, more accurate algorithm can lower aggregate outcome quality; shared base models homogenise errors. *Already shows:* accuracy–independence trade-off at the aggregate level. *Gap:* no prices, no dynamics, no adoption choice, no macro.2158. **Dou, Goldstein & Ji 2025 (NBER w34054)** — RL trading agents sustain supra-competitive outcomes through price-trigger strategies and through *homogenised learning biases*; informativeness and liquidity fall. *Already shows:* homogeneous AI biases reduce price efficiency in a Kyle-type market. *Gap:* agents are traders, not households/firms; no delegation decision; no macro prices; mechanism is learning, not a shared forecasting model.2169. **Meng & Chen 2026 (arXiv 2604.03272)** — endogenous AI adoption in a Kyle-type REE with correlated algorithmic signals; systemic-risk multiplier convex in the adoption share; supermodular adoption game with tipping to monoculture; 13F evidence. *Already shows:* **endogenous adoption + correlated signals + superlinear aggregate risk** in a financial market. This is the most dangerous overlap for H2. *Gap:* no household/firm GE (wages, *r*, capital), no explicit 1/*M* model-count comparative statics, no constrained-efficiency/Pigouvian result, no compute market, no collateral block (abstract-level reading; full text should be read before submission).21710. **Hadfield & Koh 2025 (arXiv 2509.01063)** — verbal: GE distortions from AI consumers vanish only if markets are large *and AI mistakes are zero-mean and independent*; copying the same agent induces correlated mistakes that do not wash out. *Already shows:* the premise of WP25, stated in words. *Gap:* no model, no propositions, no prices, no policy.21811. **Chaboud et al. 2014; Khandani & Lo 2011** — empirical: algorithmic strategies are more correlated than human ones; crowded quant strategies unwound together in August 2007. *Already shows:* correlated algorithmic behaviour and joint unwinding are documented facts. *Gap:* no theory of why, no welfare statement.21912. **He 2026 (SSRN 6350140); Spiro 2026 (arXiv 2605.00844); Lopez-Lira 2025; Ross et al. 2026; Hansen & Lee 2025** — the 2025–26 empirical/agentic cluster: ChatGPT acts as a belief homogeniser for retail investors; LLM errors are correlated even *across* vendors; LLM traders may behave in correlated ways; LLM agents herd less than humans in cascade games. *Already shows:* the premise has empirical support and the agentic-market experiment has been run in simple form. *Gap:* no theory linking adoption share and model count to prices and welfare; no household/firm side; and Spiro's cross-vendor correlation and Hansen–Lee's low herding are *constraints* WP25 must satisfy, not confirmations.220221### B. What is **not** new (say so in the paper)222223- "Common signals/common noise create aggregate fluctuations" — Lorenzoni 2009, Angeletos & La'O 2013, Benhabib–Wang–Wen 2015.224- "Agents over-use common information; its social value can be negative" — Morris & Shin 2002, Angeletos & Pavan 2007.225- "Information-source or information-technology choice is endogenous and can homogenise" — Veldkamp 2006, Hellwig & Veldkamp 2009, Farboodi & Veldkamp 2020, Brock & Hommes 1998 (predictor choice).226- "Private gains from similarity create inefficient common exposure; diversity is under-provided" — Stein 2009, Wagner 2011; monoculture — Kleinberg & Raghavan 2021.227- "Pecuniary externalities under incomplete markets make equilibria constrained-inefficient; Pigouvian corrections" — Greenwald–Stiglitz 1986, Lorenzoni 2008, Dávila–Korinek 2018.228- "Collateral constraints amplify shocks; margin spirals" — Kiyotaki–Moore 1997, Brunnermeier–Pedersen 2009, Geanakoplos 2010.229- "Automation raises returns to wealth and concentrates it" — Moll–Rachel–Restrepo 2022; compute as the binding constraint — Korinek–Suh 2024, Restrepo 2025.230- "Homogeneous AI/algorithmic agents reduce price efficiency and raise systemic risk; adoption can be endogenous" — Dou–Goldstein–Ji 2025, Meng–Chen 2026, Qiu–Han 2026, Park 2026.231- The verbal premise "copied agents induce correlated mistakes that do not wash out" — Hadfield & Koh 2025; policy versions in Gensler–Bailey 2020, Danielsson et al. 2022, FSB/IMF/BIS 2024.232233### C. What is genuinely new (and must be proven, not asserted)2342351. **Endogenous delegation determines the error-covariance structure of the economy.** No paper above makes the *correlation structure of forecast errors* (not just their variance) an equilibrium object chosen by households and firms through a delegation decision, with a closed-form loading on *D* and 1/*M*. Closest: Meng & Chen 2026 (traders only, adoption share only, no *M*), Hellwig & Veldkamp 2009 (correlation via complementarity).2362. **The delegation externality in a production economy.** The constrained-inefficiency result — competitive delegation is excessive for finite *M* because each delegator ignores the increase in common noise loaded on wages, *r* and asset prices — is not in the literature. Stein/Wagner have the portfolio analogue; Angeletos–Pavan the information-use analogue; nobody has the delegation version with macro prices. The Pigouvian delegation tax / model-diversity equivalence is new *if* derived with the Dávila–Korinek decomposition.2373. **The compute-market link.** A single compute rental price clearing inference demand (from delegation) and automation demand (task-based production), generating rents to algorithmic-capital owners, is new as a mechanism. It must be shown to differ from a generic rise in *r* (Moll–Rachel–Restrepo 2022) — e.g., by a distinct comparative static (compute price responds to *D*, which *r* does not).2384. **Agentic market experiments that vary *M* and *D* factorially** with LLM agents in Walrasian, CDA and bilateral designs, tied to a theory prediction. Lopez-Lira 2025, Ross et al. 2026, Qiu–Han 2026 run related experiments but without the adoption margin or the model-count treatment linked to a GE theory.239240### D. Competing explanations the referee will raise241242- **Belief homogenisation may *reduce* volatility and bubbles** (Miller 1977, Harrison–Kreps 1978, Scheinkman–Xiong 2003, Simsek 2013): fewer disagreements → less speculative volume, less leverage. WP25 must show the model-noise channel dominates the disagreement channel, or characterise when each wins.243- **Observation-based herding is not the channel** (Hansen & Lee 2025 find LLM agents herd *less*). WP25's channel is technological, so this is consistent, but the paper must not claim "LLMs herd".244- **Nominal vs effective *M*** (Spiro 2026: cross-vendor error correlation 0.77). If independent vendors share training data, 1/*M* overstates diversification; *M* must be the effective number of independent models and the policy "model diversity" may be weaker than claimed.245- **Data abundance alone lowers price informativeness** (Dugast & Foucault 2018) without any correlated-error story.246- **Learning-based collusion** (Calvano et al. 2020, Dou–Goldstein–Ji 2025, Fish et al. 2024) produces correlated AI behaviour through strategic learning, not shared forecast errors; WP25's experiments must separate the two (e.g., non-strategic price-taking agents with shared vs independent models).247- **Inelastic markets** (Gabaix–Koijen 2021): if markets were elastic, correlated flows would not move prices; the model needs limited risk-bearing capacity for model noise to be priced — state this assumption.248- **Approximate aggregation** (Krusell–Smith 1998): if delegation is wealth-dependent, the distribution may matter; the numerics must check this rather than assume it.249- **Returns to wealth** (Hubmer–Krusell–Smith 2021, Moll–Rachel–Restrepo 2022) already explain rising top wealth shares; H3 needs a distinct signature.250251### E. Verdict (referee's summary)252253The *ingredients* of the mechanism are all in the literature: common noise → fluctuations (Lorenzoni; Angeletos–La'O), over-use of common information (Morris–Shin; Angeletos–Pavan), endogenous homogenisation of information sources (Veldkamp; Hellwig–Veldkamp; Farboodi–Veldkamp), the private-gain/common-exposure externality with under-provided diversity (Stein; Wagner; Kleinberg–Raghavan), pecuniary-externality welfare analysis (Greenwald–Stiglitz; Dávila–Korinek), collateral amplification (Kiyotaki–Moore; Brunnermeier–Pedersen), and — critically — a 2026 finance paper (Meng & Chen) with endogenous AI adoption and correlated signals in a Kyle market, plus a 2025 chapter (Hadfield & Koh) that states the premise verbally. **The combination is new**: an endogenous delegation margin for households and firms that sets the error-covariance structure, a *D*/*M* loading of macro prices on model noise, a constrained-inefficiency and Pigouvian/model-diversity result in a production economy, a compute market tying inference to automation, and controlled agentic-market experiments that vary *M*. The paper is publishable as a *synthesis-with-new-theorems* contribution only if (i) the externality result is derived and shown not to collapse into Angeletos–Pavan or Wagner under relabelling, (ii) Meng & Chen 2026 and Hadfield & Koh 2025 are cited and differentiated in the introduction, (iii) *M* is treated as an effective count calibrated to measured cross-model error correlation, and (iv) the disagreement-reduction channel is confronted. Novelty gate: **passed conditionally** on (i)–(iv).254