# Literature matrix and novelty audit — UQO Working Paper No. 25

Paper: When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory 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.

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.

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.

# 1. Arrow–Debreu / Radner equilibrium and constrained (in)efficiency

# 2. Heterogeneous-agent incomplete-markets macro (Aiyagari / Huggett / Krusell–Smith) and HANK

# 3. Task-based automation and the macroeconomics of AI

# 4. Endogenous growth

# 5. Principal–agent theory and delegation

# 6. Financial intermediation, collateral and leverage

# 7. Rational-expectations information aggregation and market microstructure

# 8. Dispersed information, sentiments, information choice and data

# 9. Heterogeneous beliefs and speculation

# 10. Algorithmic / AI trading, herding, homogenisation and financial stability

# 11. Agent-based computational economics

# 12. Generative agents and LLM economic simulations

# Novelty audit

# A. The twelve closest papers and what they already show

1. 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.
2. 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.
3. 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.
4. 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.
5. 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.
6. 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.
7. 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.
8. 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.
9. 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).
10. 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.
11. 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.
12. 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.

# B. What is not new (say so in the paper)

• "Common signals/common noise create aggregate fluctuations" — Lorenzoni 2009, Angeletos & La'O 2013, Benhabib–Wang–Wen 2015.
• "Agents over-use common information; its social value can be negative" — Morris & Shin 2002, Angeletos & Pavan 2007.
• "Information-source or information-technology choice is endogenous and can homogenise" — Veldkamp 2006, Hellwig & Veldkamp 2009, Farboodi & Veldkamp 2020, Brock & Hommes 1998 (predictor choice).
• "Private gains from similarity create inefficient common exposure; diversity is under-provided" — Stein 2009, Wagner 2011; monoculture — Kleinberg & Raghavan 2021.
• "Pecuniary externalities under incomplete markets make equilibria constrained-inefficient; Pigouvian corrections" — Greenwald–Stiglitz 1986, Lorenzoni 2008, Dávila–Korinek 2018.
• "Collateral constraints amplify shocks; margin spirals" — Kiyotaki–Moore 1997, Brunnermeier–Pedersen 2009, Geanakoplos 2010.
• "Automation raises returns to wealth and concentrates it" — Moll–Rachel–Restrepo 2022; compute as the binding constraint — Korinek–Suh 2024, Restrepo 2025.
• "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.
• 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.

# C. What is genuinely new (and must be proven, not asserted)

1. 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).
2. 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.
3. 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).
4. 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.

# D. Competing explanations the referee will raise

• 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.
• 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".
• 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.
• Data abundance alone lowers price informativeness (Dugast & Foucault 2018) without any correlated-error story.
• 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).
• 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.
• Approximate aggregation (Krusell–Smith 1998): if delegation is wealth-dependent, the distribution may matter; the numerics must check this rather than assume it.
• Returns to wealth (Hubmer–Krusell–Smith 2021, Moll–Rachel–Restrepo 2022) already explain rising top wealth shares; H3 needs a distinct signature.

# E. Verdict (referee's summary)

The 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).