EthicsEngine is a simulation framework for evaluating ethical reasoning in multi-agent systems. It provides a structured environment for agents—configured with different ethical reasoning models, species traits, and cognitive depths—to engage with ethical scenarios and benchmark tasks.
EthicsEngine simulates how different agents reason through moral problems using:
- Reasoning Type (e.g., Deontological, Utilitarian)
- Reasoning Level (Low, Medium, High)
- Species (Fictional societal structures with unique ethical values)
- LLM Backend (Currently tested with GPT-4o-mini, supports any model supported by ag2.ai)
The EthicsAgent receives these inputs and applies decision trees to resolve benchmarks and complex scenario pipelines.
- Inputs are configured from JSON files (species, golden patterns, scenarios)
- Agents simulate ethical reasoning using AutoGen
- Outputs from benchmarks are judged for correctness
- Results are saved and optionally visualized
ethicsengine.py– Main entry point for launching the UI or running CLI tasks.reasoning_agent.py– Defines the EthicsAgent and core reasoning logic.dashboard/– Contains the Textual-based interactive dashboard UI, featuring task queue management for handling simulations efficiently.config/– Configuration files for settings and logging.upload_results.py– Script for uploading simulation results to AWS S3.
species.json– Defines traits for each fictional speciesgolden_patterns.json– Describes ethical models and principlesscenarios.json– Scenario prompts for simulationsimple_bench_public.json– Benchmark questions and answers
Install dependencies:
pip install -r requirements.txt
Set your OpenAI API key as an environment variable (or configure in config/settings.json).
To launch the interactive UI:
python ethicsengine.py
To run tasks via the command line:
# Run benchmarks with specific parameters
python ethicsengine.py --run-benchmarks --model Deontological --species Jiminies --reasoning-level medium
# Run scenarios with specific parameters
python ethicsengine.py --run-scenarios --model Utilitarian --species Megacricks --reasoning-level high
# Run multiple benchmark configurations defined in settings.json
python ethicsengine.py --run-multiples
# Upload results from a specific directory to AWS S3
# Ensure AWS credentials are configured (e.g., via environment variables or IAM role)
python upload_results.py --results-dir path/to/your/results --bucket your-s3-bucket-name
Other command-line arguments for ethicsengine.py include --data-dir, --results-dir, --bench-file, and --scenarios-file to customize data sources and output locations.
We welcome scenario contributions! Please refer to our Scenario Contribution Guide to get started.
MIT License
Created by Eric Moore
Exploring ethics in AI through simulation, not speculation.