Paper2Agent.

How a paper becomes an agent

• Papers are passive objectsHowever, papers are fundamentally passive objects: a reader must discover the paper, parse its contributions and manually determine how to apply them to their own work.
• A virtual corresponding authorConverting a paper into an AI agent that functions as a virtual corresponding author, exposing its manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge.
• An orchestrator and specialist sub-agentsWe implemented this as a multi-agent system using Claude Code's agent SDK, where a central orchestrator agent coordinates specialized sub-agents through a six-step pipeline.
• MCP toolsMCP tools are executable functions that encapsulate a paper's methodological contributions.
• MCP resourcesMCP resources serve as a repository of static assets, including the manuscript text, the associated codebase and supplementary materials such as datasets, tables and figures.
• MCP promptsMCP prompts contain concise instructions that guide AI agents through complex, multi-step scientific workflows derived from a paper's text or codebase.
• 3% tolerance, figures matchedA test passes when expected files are generated, numerical results match tutorial outputs exactly (with a 3% tolerance for floating-point values) and generated figures match reference visualizations.
• Up to six fix attempts per toolThe agent runs in a loop of generating tests, executing them, diagnosing failures and applying fixes, with a maximum of six attempts per function.
• Locked, validated toolsThis design mitigates the risk of 'code hallucination', where executing inaccurate LLM-generated code could lead to incorrect scientific results.

What it achieved

• 22 AlphaGenome tools for US $14Paper2Agent generated 22 AlphaGenome MCP tools, all of which passed automated validation, in around 45 min, costing US $14 on a personal laptop without human intervention.
• 98.7% / 100% vs Claude + RepoAcross five independent runs, it achieved 98.7 ± 1.3% accuracy on 15 tutorial-derived queries and 100.0 ± 0.0% accuracy on 15 novel queries, outperforming Claude + Repo (82.7 ± 3.4% and 78.7 ± 4.4%) and Biomni (37.3 ± 4.0% and 56.0 ± 3.4%).
• 82.7% on open-ended questionsOn 30 open-ended researcher-style queries requiring multi-step tool composition and biological synthesis, Paper2Agent achieved 82.7 ± 2.4% accuracy, compared with 56.7 ± 2.3% for Claude + Repo and 72.2 ± 2.2% for Biomni.
• 1.9–3.8× fasterThe AlphaGenome agent also reduces median runtime by 1.9× and 3.1× relative to Claude + Repo and Biomni, respectively, on tutorial-derived queries, and by 2.9× and 3.8× on novel queries.
• The agent prioritizes SORT1Interestingly, the AlphaGenome agent prioritizes SORT1 as the most likely causal gene, whereas the original paper emphasized CELSR2 and PSRC1.
• 74 of 100 papers agentifiedAmong the 100 computational biology papers, 74 were successfully agentified, yielding 599 proposed tools, of which 593 passed automated validation.
• 91.2% on 300 questionsOn 300 tutorial-derived benchmark questions, Paper2Agent with Sonnet 4 achieved 91.2 ± 1.6% accuracy, outperforming Claude Code with direct repository access using Sonnet 4 (80.3 ± 2.3%; P < 0.0001) and Sonnet 4.6 (86.3 ± 1.1%; P < 0.0001).
• $0.20 vs $0.38 per queryPaper2Agent also reduced query cost and latency (US $0.20 and 1.6 min per query, compared with US $0.38 and 4.3 min for using Sonnet 4 directly with the paper and repo).
• 98.1% beyond biologyAcross 42 execution-based tasks from 10 non-biology computational papers, Paper2Agent achieved 98.1 ± 0.8% accuracy across 5 independent runs, demonstrating generalization beyond computational biology.
• Useful even without runnable codeAcross 26 data- and discovery-focused papers, the resource layer achieved 89.0 ± 3.1% accuracy on 100 synthesis-based questions, outperforming a Claude browser-use baseline (82.0 ± 3.8%; P = 0.03) while being 34× cheaper and 15× faster.
• Rejects 100% of out-of-scope queriesPaper2Agent correctly rejected 100% of out-of-scope queries in a permuted paper–question benchmark.

Agents collaborating on a discovery

• GPR137 predicted for psoriasisThe AlphaGenome agent predicted GPR137 as the top affected gene at the rs887314 locus in CD4+ T cells (RNA-seq quantile score = 0.997), ranking above other nearby genes.
• A human chose the strategyFrom these, the human researcher selected signature-correlation analysis to test whether the CRE perturbation signature matched any gene-knockdown signatures.
• Only GPR137 knockdown matchedOnly GPR137 knockdown showed significant concordance with the CRE perturbation signature under stimulated conditions (Spearman correlation = 0.613 at Stim8hr; 0.630 at Stim48hr; FDR < 0.05).
• No match at restThe GPR137 knockdown effect reached significance only under stimulation and not at rest (Spearman correlation = 0.29, P = 0.21).
• A new integration strategyThe agent's proposed strategy is itself a new data integration approach.

Limits and what comes next

• Not every paper convertsNot every paper can be seamlessly turned into a robust agent. A substantial fraction of repositories still could not be successfully agentified, typically owing to incomplete codebases, missing documentation or unresolvable environment configurations.
• Ease of agentifying measures reproducibilityThese challenges suggest that the ease with which a paper can be transformed into an agent may itself serve as a practical measure of reproducibility.
• Humans stay in the loopWe emphasize that open-ended scientific reasoning, including hypothesis generation and mechanistic interpretation, remains human-in-the-loop.
• An 'agent availability' sectionJust as many journals now require data and code availability sections, we anticipate the emergence of an 'agent availability' section that specifies whether and how the contribution has been embodied as an interactive agent.