LLM-Powered Agent Simulation

The Extension

Instead of rule-based agents, we use Claude Opus as the reasoning engine:

Given:
  - Your persona (income, location, preferences, habits)
  - Real location data (competitors, distance, demographics)
  - Mathematical model outputs (Huff probability, catchment zone)
  - Company concept (what they want to open, pricing, target)

Reason as this person: Would you eat here? Why or why not?

Data Requirements

Inherits all datasets from ABM Theory, plus:

Dataset	What We Use	Link
FEHD Restaurant Licences	Agent decision context (competitor options in prompt)	View →
Ethnicity	LLM persona generation (cultural dining preferences)	View →
Household Income	Agent budget constraints in prompts	View →
Border Crossing Queues	Tourist/mainland visitor agent injection rate	View →

Why Opus, Not Sonnet

Factor	Opus	Sonnet
Persona consistency	Excellent	Good
Nuanced tradeoffs	Excellent	Good
Contradictory reasoning	Handles well	Sometimes flattens
Cost per agent	~$0.05	~$0.01

For 10 agents, total cost: ~$0.50. Trivial compared to the analysis value.

How Company Input Changes Agents

This is the critical link. Each agent receives:

1. Open Data — same for all agents (restaurants, demographics, transport)
2. Model outputs — Huff probability, catchment zone
3. Company concept — your business profile, pricing, target customer

Different business concepts get different agent reactions at the same address. The models calculate different β values. The agents reason about different tradeoffs for a wine bar versus a quick-service lunch spot.

Note

The agents don’t just say yes/no. They explain why, revealing barriers and opportunities that pure math misses.

Cost of the Full Pipeline

Component	Method	Cost
Open Data	CKAN API, FEHD XML, MTR CSV	Free
Mathematical models	Huff + Gravity + Catchment	Free
Agent simulation	10 × Claude Opus	~$0.50
Total		~$0.50

A comparable analysis from a retail consultancy: $50,000-100,000.

Roadmap

MVP (Current Implementation)

10 synthetic personas generated from demographic data for the target area. Each persona is prompted with:

• Real location data pulled from open datasets
• Mathematical model outputs (Huff probability, catchment population, site score)
• The company’s business concept, pricing, and target customer profile

Claude Opus reasons as each persona and produces structured output: visit likelihood, key objections, unmet needs, and a natural-language explanation.

Future: MirrorFish Integration

The MVP uses 10 hand-crafted synthetic personas. The next stage replaces these with MirrorFish — a population-scale agent simulation platform that generates hundreds of statistically accurate synthetic agents from demographic distributions.

	MVP	MirrorFish
Agent count	10	Hundreds
Persona source	Hand-crafted from demographics	Statistically generated populations
Demographic accuracy	Approximate	Census-calibrated distributions
Cost	~$0.50	Higher compute cost
Status	In development	Planned

MirrorFish agents would reflect the full heterogeneity of the target area: the right proportion of elderly residents, young professionals, domestic workers, tourists, and office commuters — each with internally consistent attributes drawn from real demographic data.

Implementation Notes

Implementation vs Theory

This section documents the actual implementation in functions/api/analyze.ts, including calibration decisions and bug fixes. The theoretical formula above is the target; the implementation below is what currently runs.

Current Implementation (2026-03-25)

Still a placeholder: No LLM inference occurs server-side. The function returns a structured response that the frontend can use to run inference client-side.

Prompt template: A fully-formed prompt string is generated and returned, incorporating real context from the request:

• District name and key demographics (population, median income)
• Competitor count within 800m
• Cuisine type(s), price band, floor area, target customers
• Instructions to simulate 5 named personas with step-by-step reasoning

Hardcoded persona profiles: personaProfiles array contains 5 pre-written entries — Office Worker, Local Resident, Tourist, Foodie/Expat, Budget Student — each with:

• reasoning: templated string referencing cuisine and price band
• visitLikelihood: string like “Medium-High”, “Low-Medium”

Note: These are not LLM-generated. They fill the expected output structure before real inference is wired.

Placeholder verdict: A pre-written verdict string based on competition count and median income thresholds — gives a rough summary without LLM reasoning.

requiresLLM: true: Always present in the response. Signals to the frontend that client-side inference is expected to replace the placeholders with real LLM output.

Changelog

Date	Change	Why
2026-03-25	Prompt template updated with richer context (floor area, target customers, structured task)	Initial template lacked service model and target customer context; makes LLM output more actionable
2026-03-24	Initial implementation. Placeholder from launch; LLM inference deferred to client-side.	—