This repository hosts the first-draw implementation of the adversarial multi-agent framework described in the higher-level design document (../½á¹¹¿ò¼Ü.md).
src/multiagent_sim/agents/: Agent abstractions and concrete implementations (generator, detector, user, evaluators).community/: User persona models, social graph, and experience memory helpers.core/: Simulation loop, event bus, configuration, and orchestrator.optimization/: Strategy optimizers for generator and detector agents.pipelines/: High-level workflows for data ingestion, simulation, and evaluation.scoring/: Metrics and feedback calculation logic.tools/: External tool adapters (e.g., Wikipedia search).utils/: Shared helpers (logging, identifier generation, persistence adapters).data/: Optional storage helpers and dataset interfaces.
tests/: Future unit and integration tests (currently empty).
- Create a Python 3.10+ virtual environment.
- Install the dependencies:
pip install -r requirements.txt. - Provide credentials via environment variables before running the simulation:
OPENAI_API_KEY
These keys are not bundled in the repository; set them manually when you are ready to execute the workflow. Note: Wikipedia search endpoints do not require an API key, but outbound network access must be enabled.
The src/multiagent_sim package exposes a configurable SimulationOrchestrator that wires all agents together. A minimal usage example is provided in run_simulation.py.
python run_simulation.py
This is a scaffolding project: most modules contain extensible class skeletons and TODO markers awaiting domain-specific logic, strategy optimizers, and production-grade prompts.