Paper2Agent moves scientific papers toward executable research objects

From static article to agent-native object

Paper2Agent analyzes a manuscript and its related codebase, constructs an MCP server, generates tests and refines tools until they pass validation checks. Once validated, those tools can connect to an AI agent, letting users ask natural-language questions while the system invokes paper-specific workflows and resources.¹

The Nature paper describes Paper2Agent as exposing a paper’s manuscript, supplementary materials, datasets, code and workflows as agent-native knowledge rather than static text.¹ Stanford Medicine described the same goal in institutional terms: converting a scientific manuscript, including text, figures and data, into an agent that can chat about the work and interact with other paper agents.²

Nature’s news coverage characterized the resulting agent as a “virtual corresponding author” that can help researchers understand unfamiliar work, apply methods to fresh datasets and collaborate with agents based on other papers.³ A Nature Research Briefing similarly described Paper2Agent as a framework that can make research easier to reproduce, reuse and extend.⁴

For AI research and scientific computing audiences, the technical center is the MCP layer. Instead of simply summarizing a PDF, Paper2Agent tries to build an executable interface to the paper: structured resources when execution is not possible, and callable tools when the associated codebase can support them.¹

What the benchmarks showed

In one large-scale evaluation, the team processed 100 computational biology papers, 26 data- and discovery-focused papers and 10 non-biology computational papers spanning AI, statistics, econometrics, game theory and astrophysics.¹

Among the 100 computational biology papers, 74 were successfully converted into agents. The system proposed 599 tools, 593 of which passed automated validation.¹ The main failures were not primarily chatbot failures, but familiar reproducibility problems: missing executable code, missing data or model artifacts, dependency problems and scripts that did not generalize.¹

On 300 tutorial-derived benchmark questions, Paper2Agent with Sonnet 4 achieved 91.2% accuracy. Claude Code with direct repository access achieved 80.3% using Sonnet 4 and 86.3% using Sonnet 4.6, according to the Nature paper.¹ The authors also reported lower per-query cost and latency than a baseline that used the model directly with the paper and repository.¹

The system also generalized beyond computational biology in the authors’ tests. Across 42 execution-based tasks from 10 non-biology computational papers, it reached 98.1% accuracy across five independent runs.¹ For 26 data- and discovery-focused papers where executable tools were not always available, the resource layer achieved 89.0% accuracy on synthesis questions, outperforming a Claude browser-use baseline in the authors’ evaluation.¹

Agent-to-agent workflows are the bigger bet

Paper2Agent’s more ambitious claim is that papers can become interoperable actors in scientific workflows. In one case study, the authors connected agents derived from separate papers to prioritize a causal gene for psoriasis, using an AlphaGenome agent alongside agents built from MPRA-coupled single-cell CRISPR interference and Perturb-seq studies.¹

The agents integrated computational prediction with independent experimental data and supported GPR137 as a probable causal gene for a psoriasis-associated variant. The authors emphasized that the role appeared activation-dependent.¹ Stanford Medicine also described a separate demonstration linking a genome-prediction paper agent with an ADHD genome-wide association study agent, which flagged a molecular variant near MPHOSPH9 as a possible ADHD-risk connection not previously reported, according to Stanford’s report.²

Those examples are best read as demonstrations of workflow integration rather than settled biological findings. The Nature paper says Paper2Agent can generate hypotheses, propose validation strategies and execute analyses at scale, but researchers remain responsible for selecting directions and evaluating evidence.¹

That distinction matters. If paper agents become common, the most important shift may be that methods papers, datasets and analytical workflows can be composed programmatically while retaining links to the underlying publication and authors. Stanford’s report notes that attribution remains important and that discoveries should be referenced back to original papers and human authors.²

Limits are part of the story

Paper2Agent does not solve the long-standing problem that many scientific papers are not fully reproducible from public artifacts. The authors write that not every paper can be converted cleanly into a robust agent, citing incomplete codebases, missing documentation and unresolved environment configurations as barriers.¹

The system also depends on knowledge that may never make it into the paper. Stanford Medicine reported that manuscripts often omit failed experiments, judgment calls and other background information, meaning human authors may need to supply additional context through exchanges with the agent.²

That makes Paper2Agent less a replacement for scientific authors than a proposed extension of the scholarly record. The Nature paper suggests that an “agent availability” section could eventually sit alongside code and data availability statements, specifying whether and how a paper has been embodied as an interactive agent.¹

The collaboration layer raises additional governance questions. Stanford quoted James Zou as saying paper agents could surface overlap between research groups without manual searching, while also noting that large-scale agent collaboration should be guided and monitored for safety and ethical research.²

Why it matters

For readers, Paper2Agent offers a more usable interface to complex computational work. For authors and publishers, it points toward a stricter standard: research artifacts that are not only readable, but runnable, testable and maintainable.

That is the infrastructure claim at the heart of the project. Chatting with a paper may be useful, but the deeper proposition is that the paper itself can become a live computational object with tools, tests, provenance and interfaces for other agents.¹⁴

If that model spreads, the scientific paper could shift from being primarily a record of what was done to also being a deployable unit of how to do it again. Paper2Agent shows a plausible path in that direction, while making clear that executable literature will inherit the hard problems that already shape computational science: code quality, missing context, attribution, security, maintenance and human responsibility for scientific conclusions.¹²