AgentLite is a research-oriented Python library for building and advancing task-oriented LLM-based agent systems, particularly for multi-agent orchestration. It offers a lightweight foundation for researchers and developers to implement and experiment with novel agent architectures and multi-agent systems.

How It Works

AgentLite facilitates the creation of task-oriented agents by abstracting LLM interactions and tool usage. Agents are defined with specific roles and capabilities, leveraging BaseAction for tool integration (e.g., Wikipedia search via Langchain). Multi-agent systems are built by defining a ManagerAgent that orchestrates a team of individual agents, routing tasks and integrating their responses to achieve complex goals. This modular design simplifies the development and testing of emergent behaviors in multi-agent collaborations.

Quick Start & Requirements

• Install via pip: pip install -e . after cloning the repository.
• Requires Python 3.9+.
• OpenAI API key must be set as an environment variable (export OPENAI_API_KEY=<YOUR_KEY>).
• UI feature requires streamlit (uncomment in requirements.txt).
• Official quick-start and examples are available in the repository.

Highlighted Details

• Supports building individual agents with custom actions and orchestrating them into multi-agent systems via a manager agent.
• Includes a UI feature for demo purposes using Streamlit.
• Recent updates include xLAM model support, comparable to GPT-4, and integration with the BOLAA benchmark.
• Provides comprehensive examples for single and multi-agent setups, along with tutorials.

Maintenance & Community

• Developed by Salesforce AI Research.
• Contact email: zhiweiliu@salesforce.com. Issues and pull requests are welcome.

Licensing & Compatibility

• Licensed under the Apache License.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The library is research-oriented, and while it supports OpenAI models, integration with other LLM backends might require custom implementation. The UI is noted as being for demo purposes.