LemLab is a modular framework for developing AI agents powered by large language models. Developers can define custom prompt templates, chain multi-step reasoning pipelines, integrate external tools and APIs, and configure memory backends to store conversation context. It also includes evaluation suites to benchmark agent performance on defined tasks. By providing reusable components and clear abstractions for agents, tools, and memory, LemLab accelerates experimentation, debugging, and deployment of complex LLM applications within research and production environments.

Mean-Field MARL provides a robust Python framework for implementing and evaluating mean-field multi-agent reinforcement learning algorithms. It approximates large-scale agent interactions by modeling the average effect of neighboring agents via mean-field Q-learning. The library includes environment wrappers, agent policy modules, training loops, and evaluation metrics, enabling scalable training across hundreds of agents. Built on PyTorch for GPU acceleration, it supports customizable environments like Particle World and Gridworld. Modular design allows easy extension with new algorithms, while built-in logging and Matplotlib-based visualization tools track rewards, loss curves, and mean-field distributions. Example scripts and documentation guide users through setup, experiment configuration, and result analysis, making it ideal for both research and prototyping of large-scale multi-agent systems.

NeuralABM is an open-source Python library that leverages PyTorch to integrate neural networks into agent-based modeling. Users can specify agent architectures as neural modules, define environment dynamics, and train agent behaviors using backpropagation across simulation steps. The framework supports custom reward signals, curriculum learning, and synchronous or asynchronous updates, enabling the study of emergent phenomena. With utilities for logging, visualization, and dataset export, researchers and developers can analyze agent performance, debug models, and iterate on simulation designs. NeuralABM simplifies combining reinforcement learning with ABM for applications in social science, economics, robotics, and AI-driven game NPC behaviors. It provides modular components for environment customization, supports multi-agent interactions, and offers hooks for integrating external datasets or APIs for real-world simulations. The open design fosters reproducibility and collaboration through clear experiment configuration and version control integration.

Pacman AI offers a fully functional Python-based environment and agent framework for the classic Pacman game. The project implements key reinforcement learning algorithms—Q-learning and value iteration—to allow the agent to learn optimal policies for pill collection, maze navigation, and ghost avoidance. Users can define custom reward functions and adjust hyperparameters such as learning rate, discount factor, and exploration strategy. The framework supports metric logging, performance visualization, and reproducible experiment setups. It is designed for easy extension, letting researchers and students integrate new algorithms or neural network-based learning approaches and benchmark them against baseline grid-based methods within the Pacman domain.

Multi-Agent System provides a lightweight yet powerful toolkit for designing and executing multi-agent simulations. Users can create custom Agent classes to encapsulate decision-making logic, define Environment objects to represent world states and rules, and configure a Simulation engine to orchestrate interactions. The framework supports modular components for logging, metrics collection, and basic visualization to analyze agent behaviors in cooperative or adversarial settings. It’s suitable for rapid prototyping of swarm robotics, resource allocation, and decentralized control experiments.

At its core, RxAgent-Zoo is a reactive RL framework that treats data events from environments, replay buffers, and training loops as observable streams. Users can chain operators to preprocess observations, update networks, and log metrics asynchronously. The library offers parallel environment support, configurable schedulers, and integration with popular Gym and Atari benchmarks. A plug-and-play API allows seamless swapping of agent components, facilitating reproducible research, rapid experimentation, and scalable training workflows.

ReasonChain provides a modular pipeline for constructing sequences of LLM-driven operations, allowing each step’s output to feed into the next. Users can define custom chain nodes for prompt generation, API calls to different LLM providers, conditional logic to route workflows, and aggregation functions for final outputs. The framework includes built-in debugging and logging to trace intermediate states, support for vector database lookups, and easy extension through user-defined modules. Whether solving multi-step reasoning tasks, orchestrating data transformations, or building conversational agents with memory, ReasonChain offers a transparent, reusable, and testable environment. Its design encourages experimentation with chain-of-thought strategies, making it ideal for research, prototyping, and production-ready AI solutions.

DataEnvGym delivers a collection of modular, customizable environments built on the Gym API to facilitate reinforcement learning research in data-driven domains. Researchers and engineers can select from built-in tasks like data cleaning, feature engineering, batch scheduling, and streaming analytics. The framework supports seamless integration with popular RL libraries, standardized benchmarking metrics, and logging tools to track agent performance. Users can extend or combine environments to model complex data pipelines and evaluate algorithms under realistic constraints.