AgentTorch provides a platform for building and simulating large populations of autonomous agents, enabling the study of complex societal dynamics. It targets researchers and developers needing to model phenomena at scale, such as pandemics or supply chains, offering a PyTorch-like interface for differentiable, composable, and scalable agent-based simulations.

How It Works

AgentTorch leverages a novel Large Population Model (LPM) paradigm, designed for scalability and differentiability. It allows simulations of millions of agents on GPUs, enabling gradient-based learning through stochastic dynamics and interventions. The framework supports composition with other models, including LLMs and mechanistic simulators, facilitating complex agent behaviors and parameter calibration.

Quick Start & Requirements

• Install via pip: pip install agent-torch or pip install git+https://github.com/agenttorch/agenttorch
• Requires Python >= 3.9.
• Some models may require additional dependencies (see docs/install.md).
• Official documentation: https://agenttorch.readthedocs.io/en/latest/

Highlighted Details

• Enables simulating country-sized populations in seconds on commodity hardware.
• Supports differentiating through simulations for gradient-based learning.
• Composability with LLMs for agent behavior and other simulators.
• Used for optimizing vaccine distribution and tracking global supply chains.

Maintenance & Community

• Active development with contributions from Ayush Chopra.
• Community and contribution guidelines available in the repository.

Licensing & Compatibility

• Licensed under GNU Affero General Public License v3.0 (AGPL-3.0).
• AGPL-3.0 requires network-provided services using the software to make their source code available.
• Commercial licensing options are available upon request.

Limitations & Caveats

The AGPL-3.0 license may impose significant restrictions on commercial use or integration into proprietary products without a waiver. Specific hardware requirements for certain models are detailed in separate documentation.