UQO Working Papers · WP25 · v1.0
When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Shared models turn noise that cancels into noise that does not
Capital, wages and asset prices load on model noise with weight D²/M; the private value of delegation equals its social value under constant returns, and one shared model raises order-flow correlation from 0.11 to 0.97.
Abstract
Households and firms are beginning to delegate economic decisions to autonomous agents built on a few shared foundation models. This paper studies what this does in general equilibrium. A human's decision error is idiosyncratic and averages out; an AI agent's error is common to every principal that uses the same model. Delegation therefore converts noise that cancels into noise that does not: capital accumulation, asset demand and credit load on the shared error with weight D (the delegation share) and 1/M (the effective number of independent models), and wages, interest rates and asset prices inherit this model noise.
In a task-based stochastic growth economy with a compute market, the decision problem reduces to a linear-quadratic game with dispersed information in which delegation is the purchase of a correlated signal. In a price-taking economy with constant returns the private value of delegation equals its social marginal value identically: model noise is real but not an inefficiency. Closed-form wedges, Pigouvian delegation taxes and the rate at which model diversity substitutes for them are obtained when the aggregate action has non-pecuniary effects, when price risk is uninsurable, or when prices enter collateral constraints. Delegating the consumption–savings decision has almost no aggregate bite; delegating the allocation of wealth has a large one. In a heterogeneous-agent economy the rich delegate first, better decisions concentrate wealth even with broad ownership of AI, and the distributional effect of automation is decided by who owns the compute.
In an agentic market laboratory with explicit balance sheets, margin rules, defaults, three market designs and live open-weight language-model agents under hard deterministic constraints, order-flow correlation rises with D and falls with M in every design; a shared model makes markets quieter and more correlated at once; and a false shared signal is amplified by collateral in a call market (by a factor of 2.6 at a 20% margin, halved at 50%, eliminated by model diversity) but not in a continuous auction. Language-model agents herd less than Bayesian agents on the same shared signal but load prices on it more, and their proposed actions routinely violate the constraints that deterministic code enforces. Code, data provenance, prompts and audit logs are provided.
Key findings
0.11 → 0.97
order-flow correlation, private signals to one shared model
0.43
order-flow correlation with eight independent models
0
private–social wedge of delegation under constant returns
×2.6
collateral amplification of a false shared signal, 20% margin
7.7%
wage model-noise s.d. at full adoption of one model
0.68 → 0.70
wealth Gini from the delegation channel alone
Scientific contributions
- A delegation market sets the error-covariance structure
- The delegation share D and the effective number of independent models M are sufficient statistics for how much capital, wages, interest rates and asset prices load on model noise, with closed-form loadings proportional to D²/M (Proposition 2).
- Delegation is efficient in the benchmark economy
- In a price-taking economy with constant returns and complete markets the private value of delegating equals its social marginal value identically, for every D, M and degree of strategic interaction (Proposition 3), derived symbolically before any simulation and property-tested (maximum wedge 1.9e-7 on the grid).
- Closed-form wedges and taxes under three frictions
- Non-pecuniary aggregate effects, uninsurable price risk and collateral constraints each yield a signed wedge, a Pigouvian delegation tax (4.5 times the fee at M = 1 and 0.8 at M = 8 with 50% hand-to-mouth households) and the rate 1/M at which model diversity substitutes for it.
- Which decision is delegated matters most
- Delegating consumption–savings has almost no aggregate bite (a one-log-point productivity news moves desired savings by less than 0.002), while delegating portfolio allocation gives the first adopter about 3% of consumption a year and the last 0.4%, and a wage model-noise standard deviation of 7.7% with one model against 3.2% with eight.
- The rich delegate first and ownership decides distribution
- In the Aiyagari economy the information channel alone raises the wealth Gini from 0.68 to 0.70 while adding only 3% to output; AI as a production input lowers the Gini to 0.58 under broad ownership of compute and to 0.66 under concentrated ownership.
- Agentic markets: quieter, more correlated, conditionally fragile
- Across call, continuous-auction and bilateral designs with common random numbers, order-flow correlation moves 0.11 / 0.97 / 0.43 for private signals, one model and eight; in the call market a false shared signal moves prices 3.7% unlevered and 9.8% at a 20% margin, halved at 50% and eliminated by four models, while language-model agents (Qwen2.5-7B, Mistral-7B, Llama-3.2-3B) herd less than Bayesian agents yet load prices on the shared error more.
Methodology
Methodology (5)
Symbolic derivation of the delegation game
The linear-quadratic game with dispersed information in the Angeletos–Pavan class is solved symbolically (sympy), yielding the equilibrium loadings, adoption threshold, efficiency benchmark and friction wedges of Propositions 1–7 before any numerical work.
Mapping the economy to the game
Steady state with a linearised sequence-space household and a mean-variance portfolio mapping translate the production economy into game parameters, separating the savings and portfolio channels of delegation.
Heterogeneous-agent computation
Value-function iteration with a discrete delegation choice, Howard acceleration and non-stochastic simulation, with the interest rate cleared by Brent's method and compute ownership assigned broadly or to an owner class.
Factorial agentic-market experiments
Market design × delegation share × number of models with common random numbers and bootstrap bands; stress tests with a false shared signal under unlevered, levered (20% and 50% margin) and model-diverse configurations; ledger invariants asserted at every step.
Robustness and pre-registration
Sweeps of risk aversion, persistence and the effective number of models (1 to 16); pre-registered hypotheses H1–H6 with the compute crowding-out hypothesis reported as a negative result (labour share moves by at most 0.2 percentage points).
Research data
- Three nested models
- Model A: a task-based stochastic growth economy with a compute market shared by production automation and inference, whose within-period problem is a linear-quadratic game with dispersed information. Model B: a stationary Aiyagari economy with a discrete delegation choice, a decision-quality return and broad or concentrated ownership of compute. Model C: an agentic market with balance sheets, collateralised credit, margin calls and defaults under Walrasian call, continuous double auction and bilateral bargaining, run over 20 seeds with common random numbers.
- Calibration targets
- Public series fetched 2026-10-27 with provenance and SHA-256 hashes: 23 FRED series, Fed Distributional Financial Accounts, Fernald TFP, Shiller, SEC XBRL capex, EIA, Epoch AI and Census BTOS; the AI-specific behavioural and information parameters are illustrative, recorded in a parameter ledger and swept rather than estimated.
- LLM laboratory agents
- Live open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Llama-3.2-3B, 4-bit, served locally) and a deterministic surrogate behind the same interface; every action is schema-validated and clipped by deterministic code before touching a ledger, with prompts, audit logs and a replayable cache released.
Reproducibility
- End-to-end scripts 00–50 (symbolic derivations, data download, calibration targets, Models A–C, LLM lab, figures, tables) and 30 unit, integration and property tests including the zero-wedge property and Euler's theorem
- Pre-registration (experiments/pre_registration.md), parameter ledger, model specification, limitations and an adversarial referee report in docs/
- LLM experiments replayable from results/llm_cache with full audit logs; prompts and action schemas in the appendices
- Full LaTeX source of the 71-page paper with 144 verified references; raw data re-downloadable from the manifest, not redistributed
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 71 p. |