Agent Laboratory provides an end-to-end autonomous research workflow, leveraging LLM agents to assist human researchers across literature reviews, experimentation, and report writing. It aims to automate repetitive tasks, allowing researchers to focus on ideation and critical thinking, thereby accelerating scientific discovery. The system is designed for researchers seeking to optimize productivity and explore cumulative research progress via its AgentRxiv framework.

How It Works

The system operates in three phases: Literature Review, Experimentation, and Report Writing. Specialized LLM agents collaborate, integrating tools like arXiv, Hugging Face, Python, and LaTeX. This structured workflow automates literature analysis, planning, data preparation, experimentation, and report generation, facilitating cumulative progress through the AgentRxiv framework.

Quick Start & Requirements

• Installation: Clone the repository, set up a Python 3.12 virtual environment, and install requirements with pip install -r requirements.txt.
• Prerequisites: Python 3.12. Optional pdflatex for LaTeX compilation (can be disabled with --compile-latex "false").
• Running: python ai_lab_repo.py --yaml-location "experiment_configs/MATH_agentlab.yaml"
• Documentation: Paper, Website, AgentRxiv Website

Highlighted Details

• Supports OpenAI (o1, o1-mini, o1-preview, gpt-4o, o3-mini) and DeepSeek (deepseek-chat) LLM backends.
• Introduces AgentRxiv for cumulative research progress among autonomous agents.
• Allows detailed configuration via YAML files and offers a "Co-Pilot mode".
• Supports checkpointing for progress recovery and multi-language research (with caveats).

Maintenance & Community

The project is actively developed, with recent updates introducing AgentRxiv. Contact is available via email: sschmi46@jhu.edu.

Licensing & Compatibility

Source code is licensed under the MIT License, permitting commercial use and modification.

Limitations & Caveats

While multi-language support is present, it has not been extensively studied, and users are encouraged to report issues. The effectiveness of more powerful LLMs is noted, but users must balance performance with cost and computational resources.