Paper2Agent converts research papers into working AI agents

Paper2Agent, a framework described in a new Nature paper from a Stanford-led team, converts a research paper into a working AI agent in about 45 minutes for roughly $14 on a personal laptop.

Rather than leaving a paper as a PDF, the system wraps its manuscript, code, datasets and workflows behind a Model Context Protocol (MCP) server that a chat agent such as Claude Code can call directly. "Knowledge should not be static records. It really should be dynamic and interactive," Stanford computer scientist James Zou told IEEE Spectrum.

For the paper's headline case, the team agentified AlphaGenome. Paper2Agent generated 22 validated MCP tools, and the resulting agent hit 98.7 ± 1.3% accuracy on tutorial-derived queries and 100.0 ± 0.0% on novel queries, against 82.7 ± 3.4% for Claude given direct access to the same repository and 37.3 ± 4.0% for the Biomni baseline.

At scale, the framework agentified 74 of 100 computational biology papers, with 593 of 599 generated tools passing validation. Across 300 benchmark questions drawn from those papers, the agents averaged 91.2 ± 1.6% accuracy. Zou reframes that pass rate as a signal of paper quality: "Agentification itself is a useful certificate that says, 'This work is relatively complete and well documented.'"

The paper also reports that two agents built from separate papers, talking to each other, surfaced an MPHOSPH9 variant tied to ADHD risk that Zou says had not previously been reported. It is one run, not independently replicated in the paper, and offered as a proof of concept for multi-agent scientific discovery.

Five researchers we follow shared it this week, and a live gallery of Paper2Agent-built agents is running at paper2agent.ai/live.