# Research blueprint — When AI Agents Become Economic Agents (UQO WP 25)

Written 2026-10-27, before any experiment was run. Updated only in the "Status" section.

## 1. Scientific question

What changes in general equilibrium when economic *decisions* — not only production tasks — are delegated to
autonomous AI agents? Four roles of AI are kept apart throughout: (i) production technology (compute performs
tasks), (ii) information-processing service (advisory AI: the human still decides), (iii) delegated decision-maker
(the AI agent acts within a mandate; the principal retains residual claims and liabilities), (iv) a hypothetical
independent economic principal with its own balance sheet (labelled counterfactual, never a baseline).

## 2. The mechanism we isolate (the "diversification externality of delegation")

Economic decisions are taken on forecasts of payoff-relevant states. A human decision-maker's forecast error is
*idiosyncratic*: across a continuum of households or firms it washes out by the law of large numbers, so it costs
only the person who makes it. A decision delegated to an AI agent built on foundation model *m* carries model
*m*'s error, which is *common* to every principal that uses model *m*. Delegation therefore has two faces:

* **private face** — it raises forecast precision (the model is better informed than the median human) and
  saves supervision time, at a fee that is priced in the compute market;
* **aggregate face** — it replaces a noise that averages out with a noise that does not. Aggregate capital
  accumulation, asset demand and credit all start to load on the shared model error with weight
  `D` (delegation share) and `1/M` (number of independent foundation models).

Two results organise the paper.

1. **Efficiency benchmark (Proposition 3).** In a price-taking, complete-markets economy with constant returns,
   the private value of delegating equals its social marginal value *for every* `D`, `M` and degree of strategic
   substitutability. Aggregate "model noise" is real — wages, interest rates and investment fluctuate with the
   foundation model's mistakes — but it is **not** an inefficiency: delegators dampen their response to the shared
   signal exactly as a planner would (strategic substitutability in capital accumulation), and the pecuniary
   effects on others are zero-sum by Euler's theorem. This was derived symbolically (`scripts/00_lq_derivations.py`)
   before any numerical work; it is the honest starting point and the first thing a referee should see.
2. **Where delegation becomes a policy problem (Propositions 4–7).** The wedge between private and social value
   is non-zero exactly when (a) the aggregate action has non-pecuniary effects (`U_xX + U_XX ≠ 0`, e.g. knowledge
   spillovers, market power), (b) price risk created by model noise is uninsurable for non-participants
   (hand-to-mouth workers, heterogeneous ownership of compute), (c) prices enter collateral constraints (margin
   spirals; Dávila–Korinek-type pecuniary externality), or (d) execution happens in markets with finite depth
   (price impact of correlated orders). We derive the sign and size of each wedge, the corresponding Pigouvian
   delegation/compute tax, and show that model diversity `M` is a substitute for the tax in cases (b)–(d) but
   irrelevant in the benchmark.

The compute market ties the decision channel to the production channel: inference demand from delegation and
task automation compete for the same compute stock. In the short run (fixed compute) better decision-AI *raises*
the compute rental, crowds out production automation and can *raise* the labour share; in the long run
(elastic compute at user cost) both expand. Compute scarcity rents flow to the owners of algorithmic capital;
concentrated ownership raises wealth concentration even when aggregate output rises.

## 3. Three nested models

| Model | Equilibrium concept | What it answers | Solution |
|---|---|---|---|
| A. Tractable baseline | Linear-quadratic recursive competitive equilibrium with dispersed information (Angeletos–Pavan class), around the deterministic steady state of a task-based stochastic growth economy with a compute market | Props 1–5: delegation threshold, aggregate model noise, efficiency benchmark, inefficiency wedges, compute crowding-out | closed forms (sympy) + numerical cross-checks |
| B. Heterogeneous agents | Stationary recursive competitive equilibrium (Aiyagari class) with idiosyncratic earnings risk, borrowing limit, idiosyncratic return noise (human decision mistakes), delegation fee, task automation, compute ownership (broad vs concentrated) | H1, H3, H6 in the cross-section: who delegates, who owns, Gini / top shares / labour share / r / welfare by percentile; decomposition of productivity, displacement, information, ownership, borrowing effects | EGM/VFI (numba), bisection on r |
| C. Agentic asset market and banking | Bounded-rationality simulated outcome (temporary trading equilibrium per period) with explicit balance sheets, margin rules, defaults and loss allocation; Walrasian call vs continuous double auction vs bilateral bargaining | H2, H4, H5: common-model herding, depth under stress, collateral amplification, leverage limits; LLM agents vs rational/rule controls | discrete-event simulation; LLM via local OpenAI-compatible servers |

## 4. Hypotheses (pre-registered in `experiments/pre_registration.md`)

H1 adoption threshold · H2 common-model trade correlation and depth · H3 ownership concentration and wealth
concentration · H4 credit/collateral feedback · H5 diversity/leverage/ownership policies · H6 compute scarcity rents.
Each hypothesis has a mechanism, a prediction, a falsification condition, an isolation experiment and a limitation.

## 5. Calibration and validation

`docs/parameter_ledger.csv` (one row per parameter, source, moment, range). Macro moments from FRED/BEA/BLS,
wealth shares from the Fed Distributional Financial Accounts, AI parameters from Epoch AI / Census BTOS /
provider price lists, all fetched on 2026-10-27 and logged in `data/raw/MANIFEST.json`. The AI-specific
behavioural parameters (signal precisions, supervision time, delegation disutility) have no accepted empirical
counterpart: they are **illustrative** and swept over wide ranges. Validation is limited to the zero-AI limit
(standard RBC/Aiyagari moments) and to qualitative facts about algorithmic trading correlation.

## 6. LLM laboratory

Local open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Llama-3.2-3B, 4-bit, `mlx_lm.server`) on MacLustr
nodes; optional BYOK providers through environment variables only. Roles: household, investor, firm manager,
lender, market maker, AI portfolio manager, procurement agent, regulator. JSON-schema validated actions, hard
constraints enforced in deterministic code, full audit log (model, prompt hash, context hash, temperature, seed,
response, latency, tokens), content-addressed cache. Controls: Bayesian mean-variance agent, rules, single LLM,
heterogeneous LLMs, shared LLM with varied prompts. Simulated LLM behaviour is never interpreted as human
preference evidence.

## 7. Gates and deliverables

G1 novelty → `docs/literature_matrix.md`; G2 theory → `docs/model_specification.md` + Props 1–3 with proofs;
G3 numerics → `tests/` (accounting identities, zero-AI limit, LQ closed forms vs Monte Carlo, Aiyagari vs known
case); G4 agents → schema + constraint tests + audit logs; G5 science → factorial experiments with CRN and CIs;
G6 writing → `paper/main.pdf`, `docs/referee_report.md`, README, spbgit `wp25_uqo.git`.

## 8. Status

* 2026-10-27 — blueprint written; LQ derivations done (wedge = 0 in the benchmark, confirmed by direct
  expected-utility computation); local inference servers up; literature and calibration audits running.