Twenty-two tools, 45 minutes, $14, and no human in the loop. That is what Paper2Agent produced from a single genomics paper.
The Stanford system reads a study and converts its methods into software an agent can call. What comes out is a set of MCP servers: callable tools for the paper’s methods, resources holding the manuscript, code links, datasets and figures, and prompts that encode multi-step procedures such as the right order for single-cell preprocessing. Six steps run under a central orchestrator that dispatches specialised sub-agents.
Then comes the gate. Files must appear where expected. Numbers must land within 3 percent. Figures are compared to reference images by perceptual hash and must fall under a Hamming distance of 20. Six attempts are allowed per function, after which a failing tool is cut.
Replication ran on Claude Sonnet 4. The AlphaGenome build passed validation on all 22 tools untouched. A second agent for Scanpy returned 7 tools for $13, and across four public datasets its cell counts, gene counts and top marker genes lined up with what human researchers reported. A TISSUE agent repeated the exercise on spatial transcriptomics data.
Two human experts graded five runs and agreed 96.7 percent of the time. Median runtime landed 1.9x ahead of Claude Code with repository access and 3.1x ahead of Biomni, and the lead held after the baseline moved up to Claude Opus 4.6.
Failure handling was tested too. Out-of-scope queries were rejected across the board in a permuted benchmark, and the system recovered from broken dependencies, wrong file paths, typos and retired interfaces. The code carries an MIT licence and installs as a Claude Code or Codex skill.
