An AI Agent That Actually Runs the Experiments It Writes

What it does

NanoResearch is an end-to-end autonomous research engine that shoves a raw topic through a nine-stage pipeline: ideation, planning, setup, coding, execution, analysis, figure generation, writing, and review. It searches literature via OpenAlex and Semantic Scholar, drafts runnable experiment code, then dispatches training jobs to local GPUs or SLURM clusters. The resulting metrics and logs feed directly into automatically generated charts and a complete LaTeX paper, and the system retries failed runs after analyzing its own error logs.

The interesting bit

The project treats the LLM as a lab manager rather than just a ghostwriter. By routing each pipeline stage to distinct models and temperatures—warmer for brainstorming, frigid for code—it keeps hallucinations out of the training scripts while letting the literature review stay creative. The TUI and Feishu bot integrations make it clear the authors expect you to fire this off and check back after dinner to find a finished paper waiting.

Key highlights

• Nine-stage deep pipeline with resumable checkpoints between every stage
• Real GPU execution on local hardware or SLURM clusters, including automatic error-log analysis and retry logic
• Figures and tables generated from actual training logs rather than LLM fabrications
• Per-stage model routing with distinct temperature and token budgets for ideation, coding, writing, and review
• Exports a full reproducibility bundle: paper.tex, paper.pdf, figures, code, data, and references.bib

Verdict

Academic researchers and ML engineers who need rapid, reproducible prototypes will find this a useful automation backbone. If you lack API credits or GPU access, or if your work involves wet-lab pipettes instead of PyTorch, this is not your tool.