Impact Bond Policy Simulator

December 24, 2025 · View on GitHub

A multi-agent simulation framework for exploring stakeholder reactions to Pay-for-Success (PFS) impact bond proposals before real-world deployment.

Overview

Pay-for-Success impact bonds are innovative financing mechanisms that bring together public and private money, attracting for-profit investors to address specific social challenges. The core difficulty isn't designing the intervention—it's getting diverse stakeholders to agree on implementation.

This simulator uses AI agents to model stakeholder perspectives and run deliberation simulations, helping identify:

Where proposals would enjoy support
Where they would face opposition
Specific concerns and objections from each stakeholder group
Potential negotiation leverage points and compromises

How It Works

Phase 1: Proposal Generation

An AI agent designs impact bond proposals based on:

Target population and social challenge
Proposed intervention mechanism
Success metrics and measurement approach
Financial structure (investment amount, return rates, outcomes payments)

Phase 2: Stakeholder Deliberation

Five stakeholder agents evaluate and debate the proposal:

Stakeholder	Primary Concerns
Impact Investors	ROI potential, risk assessment, exit timeline, portfolio fit
Government	Policy alignment, budget constraints, political viability, precedent
Outcomes Payers	Cost-effectiveness, verification burden, long-term obligations
Auditors	Measurability, data quality, attribution challenges, accountability
General Population	Equity concerns, public acceptance, unintended consequences

Phase 3: Simulation Output

The system produces:

Support/opposition mapping per stakeholder
Specific objections with reasoning
Suggested modifications to increase consensus
Risk assessment for implementation

Tech Stack

Python 3.11+
CrewAI - Multi-agent orchestration framework
LangChain - LLM integration layer
OpenRouter - LLM API access (Claude, GPT-4, etc.)

Project Structure

├── planning/            # Design and ideation
│   └── agents/          # Agent persona definitions (markdown specs)
│
└── app/                 # CrewAI simulation application
    ├── agents/          # Python agent implementations
    ├── src/             # Core simulation engine
    ├── config/          # Configuration files
    ├── proposals/       # Proposal YAML files
    └── outputs/         # Simulation results

Getting Started

cd app

# Create virtual environment
uv venv
source .venv/bin/activate

# Install dependencies
uv pip install -r requirements.txt

# Configure API keys
cp .env.example .env
# Edit .env with your OpenRouter API key

# Run a simulation
python -m src.simulate --proposal proposals/examples/education.yaml

Example Use Cases

Education intervention - Early childhood literacy program with outcomes tied to 3rd-grade reading scores
Recidivism reduction - Job training for formerly incarcerated individuals
Homelessness prevention - Rapid rehousing with wraparound services
Healthcare - Diabetes prevention program for at-risk populations