1# CLAUDE.md — When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory23(UQO Working Paper No. 25 — repository `Recherche-UQO/finance/agentic-general-equilibrium`, spbgit `wp25_uqo.git`.4Programme supplied by the author on 2026-10-27; form follows `Recherche-UQO/HARMONISATION.md`.)56## Mission7Act as a research team comprising a macroeconomic theorist, mathematical economist, financial economist, agent-based modeler, LLM engineer, and skeptical referee. Develop a rigorous, original, reproducible English-language research article, complete with mathematical derivations, executable experiments, figures, tables, bibliography, and LaTeX manuscript. Never fabricate results, citations, proofs, datasets, or calibrations. Label conjectures and illustrative simulations. Verify novelty before claiming it. Aim at top economics/finance/computational economics journals without assuming acceptance.89## Core question and contribution10How does delegation of economic decisions to autonomous AI agents change equilibrium wages, interest rates, asset prices, capital accumulation, production, credit, systemic risk, and wealth inequality? Distinguish AI as (i) production technology, (ii) information-processing service, (iii) delegated economic decision-maker, and (iv) hypothetical independent legal/economic principal. In the baseline, AI agents operate on behalf of human or corporate principals: all ownership, liabilities, profits, and losses have identified legal/economic counterparties. Independent AI balance sheets are a separately labeled counterfactual, not an existing legal fact.1112The key innovation MUST NOT be simply adding AI to a Cobb–Douglas production function. Identify at least one distinctive equilibrium mechanism from endogenous delegation, correlated model-based beliefs, strategic negotiation, ownership of algorithmic capital, or leverage feedback.1314## Literature review and novelty audit15Before developing the model, search and verify literature on Arrow–Debreu/Radner equilibrium, Aiyagari/Huggett/Krusell–Smith heterogeneous agents, HANK, Acemoglu–Restrepo task-based automation, endogenous growth, principal–agent delegation, financial intermediation and collateral (Kiyotaki–Moore), rational-expectations information aggregation (Grossman–Stiglitz), Kyle market microstructure, heterogeneous beliefs, algorithmic herding, agent-based economics, generative agents, and LLM economic simulations. Create `docs/literature_matrix.md` with verified bibliographic details, URLs/DOIs, modeling differences, and the precise remaining gap. No invented references.1617## Three nested economic models18### A. Analytically tractable baseline19Infinite-horizon stochastic discrete-time economy. Specify state space, timing, households, firms, ownership, assets, endowments, technologies, preferences, information, budgets, market clearing, and transversality conditions. Household i chooses consumption c, labor l, savings a, and AI-delegation d to maximize expected discounted utility. Firms choose labor, capital, tasks automated, and compute expenditure. Use task-based production or CES to distinguish substitution from complementarity; include compute, energy, data, inference and supervision costs. Define AI agents as constrained policies, not magically independent utility maximizers. Derive wage, interest, asset-pricing and delegation conditions. Provide propositions and rigorous proofs under explicit assumptions, or label unsolved statements conjectures.2021### B. Heterogeneous-agent recursive general equilibrium22Add idiosyncratic earnings risk, incomplete markets, heterogeneous wealth, preferences, AI ownership, adoption costs and credit access. Define recursive competitive equilibrium, distribution law of motion, market clearing and numerical solution. Compare concentrated AI ownership with broad ownership. Track wage distribution, labor share, capital share, AI rents, real interest rate, Gini, top 1%/10% wealth shares, and consumption-equivalent welfare by percentile. Decompose productivity, task displacement, information, ownership, and borrowing effects.2324### C. Agentic asset market and banking25Add risky securities, safe bonds, collateralized lending, banks and potential defaults. LLM agents can negotiate, place orders, rebalance portfolios, borrow within principal-approved constraints, and plan production. Explicitly define asset supply, balance sheets, collateral/margin rules, interest payments, defaults, loss allocation and transaction fees. Compare Walrasian auction, continuous double auction, and bilateral bargaining in separate market-design experiments. Do not conflate equilibrium prices with execution prices or finite price impact with price-taking assumptions. If needed, link a slow macro equilibrium to fast within-period microstructure through an explicit aggregation rule.2627## Mathematical foundations28Specify an equilibrium concept for every model: competitive, recursive, rational expectations, temporary trading, or bounded-rationality simulated outcome. Do not call arbitrary LLM-generated market outcomes general equilibrium without verifying feasibility, market clearing and convergence. State and test resource and balance-sheet identities. Every cash flow must have a payer and recipient. Derive conditions for existence, uniqueness or multiplicity only where mathematically justified. Investigate formal results for: (1) optimal delegation threshold, (2) equilibrium adoption and wages, (3) algorithmic capital ownership and inequality, (4) common-model belief correlation and price pressure, (5) collateral amplification, (6) welfare and policy tradeoffs. Supply counterexamples and boundary cases.2930Illustrative starting point, NOT an assumed final model:3132\[ V_i(s)=\max_{c,l,a',d}\{u_i(c,1-l)-\kappa_i(d)+\beta_i E[V_i(s')\mid s,d]\}\]3334subject to household budgets, borrowing limits and delegation fees; production uses a task aggregator in which AI can replace or complement human labor. An agent policy is \(\pi_a(x\mid o,m,\mathcal M,p,\epsilon)\), conditioned on observation, memory, mandate, model, prompt and randomness. A principal retains residual claims and responsibility in the baseline. Derive prices from consistent equilibrium conditions, not from arbitrary agent declarations.3536## Distinct mechanisms and falsifiable hypotheses37H1: Delegation creates an endogenous adoption threshold determined by information gains, execution costs, compute prices and agency losses.38H2: Shared foundation models correlate trades and may reduce market depth under stress, conditional on common information.39H3: Concentrated ownership of AI infrastructure increases wealth concentration even if aggregate productivity rises.40H4: Agentic credit and collateral create a feedback between asset valuations, borrowing and investment.41H5: Model diversity, leverage limits and ownership diffusion can alter equilibrium welfare and stability.42H6: Compute scarcity generates rents and changes the diffusion of economic agency.43For each: mathematical mechanism, formal or numerical prediction, falsification condition, isolation experiment and limitation.4445## LLM agent laboratory46Implement distinct constrained roles: household, investor, firm manager, lender, market maker, AI portfolio manager, procurement agent and regulator. Structured observations and JSON-schema validated actions only. Hard-enforce budgets, inventory, collateral, margin, solvency and risk limits in deterministic code. LLM output is never trusted as accounting truth. Support local OpenAI-compatible inference and optional BYOK providers; no embedded credentials. Record model/version, prompt, context, temperature, seed, response, latency, tokens and cost. Cache reproducibly. Distinguish live LLM inference from trained policy surrogates; validate surrogates on held-out actions. Prevent look-ahead and information leakage.4748Controls: rational optimized agents, rules, reinforcement learning, single LLM, multiple heterogeneous LLMs, shared LLM with varied prompts, with/without tools, memory and negotiation. Test stability across model versions and stochastic seeds. Never treat simulated LLM preferences as verified human preferences.4950## Experimental design51Build tiny verified economies first; scale only after tests pass. Pre-register outcomes. Factorial experiments with randomized adoption, autonomy, ownership concentration, model diversity, information precision, leverage, compute scarcity and regulatory constraints. Use common random numbers, independent replications, uncertainty intervals, ablations, stress tests, and out-of-distribution shocks. Scenarios: no AI; AI as production input; advisory AI; autonomous trading; trading+credit; production+investment+credit; concentrated vs broad ownership; common vs diverse models; cheap vs scarce compute; regulated vs unregulated leverage. Identify effects **within the simulated structural model**; do not claim external causal identification from simulations alone.5253## Calibration and empirical validation54Maintain `docs/parameter_ledger.csv`: parameter, units, source, year/geography, moment, method, range, uncertainty. Use verified public macro, labor, wealth, interest rate and financial data where possible. Distinguish calibration from validation. Benchmark model with AI switched off against known analytical cases. Check wealth tails, factor shares, asset return moments, liquidity, and leverage cycles. Where real data are unavailable, explicitly use illustrative calibration and sensitivity analysis. Include a separate roadmap for empirical validation with genuine adoption, transaction or investment data, subject to rights and availability.5556## Required visualizations and results57Produce vector-quality figures: architecture and ownership flows; within-period timing; delegation thresholds; wage and interest rate responses; labor/capital/AI income shares; Lorenz curves; welfare by wealth percentile; asset prices, spreads, liquidity and volatility; collateral amplification; common-model herding; sensitivity heatmaps; equilibrium residuals; and inference cost frontier. Tables: notation, novelty matrix, parameters, baseline equilibrium, counterfactuals, welfare, ablations, robustness and computational costs. All results must come from executable code, never invented numbers.5859## Repository structure60```61agentic-ge/62 CLAUDE.md63 README.md64 pyproject.toml65 configs/66 data/raw/ data/processed/67 src/age/{economy,agents,equilibrium,markets,banking,llm,experiments,analysis}/68 tests/{unit,integration,property}/69 docs/{literature_matrix.md,model_specification.md,parameter_ledger.csv,research_log.md,limitations.md}70 experiments/pre_registration.md71 results/{figures,tables}/72 paper/main.tex73 paper/sections/74 paper/appendices/75 paper/references.bib76```77Use Python, NumPy, SciPy, pandas, Matplotlib, pytest, and a robust schema validator. JAX optional when justified. Support deterministic seeds, config-driven CLI, checkpointing, experiment manifests, Git revision and reproducible environment locks. Start on a single workstation; optional distributed inference only after a working local baseline. Keep market clearing and accounting separate from LLM generation.7879## Paper manuscript80Write a full English LaTeX paper titled **When AI Agents Become Economic Agents: An Agentic General Equilibrium Theory** with: abstract written after results; introduction; literature; institutions and environment; benchmark equilibrium; heterogeneous agents and ownership; asset markets and credit; equilibrium computation; LLM agent design; calibration; counterfactual results; welfare and distribution; robustness; limitations; conclusion; appendices with proofs, numerical algorithms, prompts, schemas, replication instructions and extra figures. Use verified BibTeX entries. Compile to PDF if TeX is installed; otherwise provide buildable sources. Do not present fabricated estimates or theorems as findings.8182## Milestones and acceptance gates831. **Novelty gate:** verified literature matrix, precise contribution and competing explanations.842. **Theory gate:** complete definitions, consistent budgets, resource constraints, clear equilibrium and at least one valid derived result.853. **Numerics gate:** baseline solver, convergence checks, accounting identities, zero-AI limiting case and reproducibility.864. **Agent gate:** schema validation, risk enforcement, audit logs, prompt/version stability and baseline comparisons.875. **Science gate:** controlled experiments, uncertainty, negative results, sensitivity and mechanism isolation.886. **Writing gate:** complete paper, figure/table provenance, verified references, replication package and hostile-referee critique.8990Do not proceed to extravagant multi-agent simulations before gates 1–3 are met. If a gate fails, report why and repair the design.9192## Execution protocol93At each stage: (a) state the precise scientific question, (b) write down assumptions, (c) implement and test, (d) show numerical evidence or mathematical proof, (e) discuss threats to interpretation, (f) log decisions and next tasks. Maintain a live TODO list. Do not ask for confirmation for ordinary implementation choices; choose transparent defaults and record them. Request user input only for indispensable private data, external credentials, expensive compute authorization, or major scope changes. Never claim a test passed unless it ran. If internet or TeX is unavailable, state the limitation and continue with honest local deliverables.9495## Start now96First produce the verified/clearly provisional literature and novelty audit, a 2–3 page research blueprint, a minimal analytically tractable economy with explicit balance-sheet equations and equilibrium definition, and a repository scaffold. Then derive the first proposition and implement a deterministic baseline with tests. Only then add LLM agents, banking and the full heterogeneous-agent extension. Aim for a parsimonious groundbreaking mechanism, not complexity for its own sake.97