Claude Code with a PhD: agent that runs experiments and writes papers

What it does AutoSci is a Claude Code-based agentic system that tries to automate the entire scientific research lifecycle. It ingests papers, generates research ideas, designs and executes experiments locally or remotely, drafts manuscripts, and even generates conference posters from LaTeX. Everything is tied to a persistent wiki memory that the authors claim compounds across projects.

The interesting bit The system does not merely chat about science—it actually deploys code to local or remote environments via /exp-run, evaluates results with /exp-eval, and can turn a drafted paper into a print-ready HTML poster with /poster. The memory-centric angle means it ranks new arXiv papers and conference proceedings against your existing wiki, surfacing only the dozen or so papers that actually matter for your direction.

Key highlights

• Persistent wiki memory that spans projects, used to rank literature recommendations and ground new ideas in prior work.
• End-to-end experiment pipeline: from pilot studies (/exp-pilot-run) through full design, deployment, and evaluation.
• Conference poster generation (/poster) that extracts figures and tables from LaTeX and produces a 1400×900 HTML layout.
• Daily arXiv digests and venue-specific discovery (/discover) filtered against your existing research graph.
• Knowledge graph visualization via local web server or Obsidian canvas export.

Caveats

• The project is explicitly labeled an “internal beta,” and the paper branch is a frozen research snapshot with capabilities still being implemented.
• It is tightly coupled to Claude Code; you are essentially buying into Anthropic’s ecosystem.
• The README lists papers produced “end-to-end” but does not clarify how much human intervention each stage required.

Verdict AutoSci is worth a look if you are a researcher who already lives in Claude Code and wants structured memory across long projects. If you are not doing experimental computer science or are wary of beta-grade automation touching your compute environments, steer clear.