An LLM Research Team That Handles Lit Reviews and LaTeX

What it does Agent Laboratory is an end-to-end autonomous research workflow that assists human researchers in implementing their ideas. It fields specialized LLM agents across three phases—literature review, experimentation, and report writing—integrating tools like arXiv, Hugging Face, Python, and LaTeX to search papers, run code, and compile PDFs. A companion framework called AgentRxiv also lets these agents upload, retrieve, and build on each other’s prior work.

The interesting bit The project explicitly positions itself as a complement to human creativity, not a replacement. You still do the ideation; the agents handle repetitive coding and documentation. The authors also stress that you must write extensive “notes” telling the agents exactly what compute resources they have, which APIs to use, and what experiments to run—making this more of a very literal research assistant than a self-directed scientist.

Key highlights

• Three-phase pipeline: literature review → experimentation → report generation
• Supports OpenAI (o1, gpt-4o, o3-mini) and DeepSeek (deepseek-v3) models
• AgentRxiv allows agents to collaboratively build on prior agent-generated research
• Checkpointing via state_saves so failed subtasks don’t kill progress
• Optional copilot mode for human-in-the-loop guidance

Caveats

• The authors explicitly state there is “a lot of room to improve this codebase”
• Non-English support exists but has not been extensively tested
• Currently limited to OpenAI and DeepSeek backends unless you submit a PR

Verdict Worth exploring if you want to automate grunt work in your research pipeline and don’t mind babysitting LLMs with detailed instructions. If you’re looking for a fully autonomous scientist that generates breakthrough ideas from a one-line prompt, this isn’t it.