AgentLaboratory

Agent Laboratory is a Python tool that automates parts of the academic research process using AI language models. Given a research idea, it runs through three phases on its own: first it reviews relevant papers on arXiv, then it designs and runs experiments with code, and finally it writes a report of the findings in LaTeX. The goal is to handle the repetitive, time-consuming parts of research so that a human researcher can focus on the ideas themselves.

The workflow uses multiple AI agents, each focused on a specific role within those three phases. They work together and can use external tools including arXiv for papers, Hugging Face for datasets, Python for running experiments, and LaTeX for producing written output. A human can stay involved throughout through a copilot mode, where the system pauses to ask for feedback and approval at key points rather than running fully on its own.

The system supports several AI models as its backbone, including OpenAI's GPT-4o and reasoning-focused models, as well as DeepSeek. You choose which model to use when launching the tool. More capable models generally produce better research outputs but cost more to run.

Setup involves cloning the repository, creating a Python environment, installing dependencies, and optionally installing a LaTeX compiler so the system can produce PDF reports. You configure experiments through YAML files and can add notes to guide the agents on what to focus on, what compute resources are available, and what style of output you want. Progress is saved to disk automatically, so you can resume a run that was interrupted.

The project also introduced a companion framework called AgentRxiv, where multiple autonomous research agents can share and build on each other's work across separate research sessions, allowing findings to accumulate over time.