Paper2Agent: How Research Papers Become AI Agents

Paper2Agent is a framework that turns a research paper and its supporting code, data and workflows into an interactive AI agent. Published in Nature on 16 September 2026, it builds a Model Context Protocol (MCP) server, validates executable tools against reference results and lets researchers query or apply a paper’s methods in natural language. The promise is not that an LLM replaces scientific judgment, but that a tested interface can make technical knowledge easier to reuse.

What Paper2Agent builds

A paper may describe a method while its code, data and runtime requirements remain difficult to use. Readers often have to locate a repository, install dependencies, configure credentials, learn input formats and reproduce the correct sequence of operations.

Paper2Agent treats these materials as one research artefact with two layers:

Paper2MCP: a server exposing executable tools, static resources and workflow prompts.
The paper agent: an AI interface that connects to the server and translates natural-language requests into tool calls.

This is more than retrieval-augmented generation over the manuscript. The research team describes a system that can represent the manuscript, supplementary material, codebase, datasets, examples and analysis workflows. A user can ask what a paper says, request a demonstration, regenerate a figure or apply a method to new data—if the implementation and inputs support that task.

The approach fits SignorCrypto’s MCP vs A2A guide: MCP connects an agent to tools and context, while other protocols can coordinate agents with one another.

How the paper-to-agent workflow works

1. Extract the research artefact

The system identifies the paper, associated code and files needed to run the method, then prepares an environment and records requirements. The result still depends on what the authors published: missing code or data limits what can be automated.

2. Turn methods into MCP tools

An extraction agent translates core operations into callable functions. In MCP terms, a server can expose tools for execution, resources for documents and data, and prompts for reusable workflows. The MCP specification defines these primitives and warns that tools can enable arbitrary data access or code execution.

3. Test against the reference

A testing agent runs generated tools against examples from the original work. Reported checks compare expected files, numerical outputs within tolerance and figures against reference results. Tools that repeatedly fail validation are excluded.

This is the key design choice: “the generated interface passed a defined test” is a stronger claim than “the model wrote plausible code.” It does not prove every future query is correct, but it creates a traceable baseline. The principle also appears in context engineering for AI agents: reliability depends on tools, runtime data, constraints and feedback loops around the model—not on the prompt alone.

4. Connect the server to an agent

The MCP server can be connected locally or remotely to a compatible AI client. The user asks for a workflow in ordinary language while the agent plans and invokes the server’s tools. Multiple paper MCPs can be connected to the same client, allowing methods from related publications to work together.

What the Nature case study shows

The team tested Paper2Agent on several scientific codebases. The most detailed example is an agent built from AlphaGenome, an AI model for predicting effects of DNA variants across regulatory processes.

Reported result	Meaning
22 AlphaGenome MCP tools	Coverage of functions including variant scoring and visualisation.
About 45 minutes	One-time conversion without human intervention on a personal laptop.
US$14	Approximate computing cost reported for that conversion.
98.7% ± 1.3%	Accuracy on 15 tutorial-derived queries across five runs.
100.0% ± 0.0%	Accuracy on 15 novel queries across five runs.
82.7% ± 2.4%	Accuracy on 30 open-ended researcher-style queries.

These are the authors’ results, not a general benchmark for every paper or domain. In the reported comparison, the AlphaGenome agent outperformed a general coding agent with repository access and Biomni on the tested tasks, and reduced median query time in those experiments.

The case study also produced a warning. For a genetic variant associated with LDL cholesterol, the generated agent prioritised SORT1, whereas the original AlphaGenome paper had emphasised CELSR2 and PSRC1. The source data supported evidence for all of these nearby genes. An agent can help re-examine a conclusion; it does not turn an ambiguous biological association into a settled fact.

The Paper2Agent repository documents quick-start instructions, supported coding-agent hosts, example conversions for TISSUE and Scanpy, and connectable paper MCP servers.

Why this changes research workflows

The immediate benefit is lower friction between reading a method and trying it. A biologist may ask for a variant analysis without learning every API parameter. A computational researcher can inspect a tool’s source reference and reproduce a validated operation instead of relying on generated code.

The deeper shift is in the unit of knowledge. A paper can become a versioned, executable interface with:

A knowledge layer: manuscript text, supplementary material and metadata.
An execution layer: tested functions wrapping the method.
A workflow layer: prompts and tool sequences for multi-step tasks.
An evidence layer: source references, test cases, outputs and limitations.

The lesson also applies beyond academic publishing. Companies face the same problem when procedures are trapped in documents, repositories and specialist memory. The transferable pattern is not “turn every document into a chatbot,” but “package knowledge with controlled tools, permissions, tests and observable outputs.” Teams evaluating that pattern should track agent observability metrics, including tool calls, latency, cost, failures and audit trails.

Limits and security questions

Paper2Agent reduces setup work; it does not remove scientific or operational responsibility.

Finite validation: passing benchmark cases does not establish correctness for every dataset, parameter or domain.
Source dependency: undocumented assumptions, unavailable data or unstable dependencies constrain the agent.
Consequential actions: a natural-language request may trigger code execution, external API calls or access to sensitive data.
Remote-server risk: the MCP specification calls for consent, clear authorisation, access controls and caution around tools and data transmission.
Human interpretation: experts must assess data quality, assumptions, causal claims and whether an output is fit for a decision.

A responsible deployment should begin with read-only or sandboxed tasks, pin environments and datasets, record tool inputs and outputs, require approval for external side effects and keep source references visible. Treat the agent as software with provenance, not as an authoritative author.

A practical adoption checklist

Choose one method with a maintained repository and clear example outputs.
List inputs, credentials, datasets and external services.
Define acceptance tests before generating the agent.
Expose the smallest useful set of MCP tools.
Compare outputs with the source implementation on known cases.
Record method version, data version, tool calls and model used.
Test malformed inputs, prompt injection and unauthorised access.
Keep human approval for high-impact actions.

Frequently asked questions

Is Paper2Agent just an AI summary of a paper?

No. Summarisation operates mainly on text. Paper2Agent also attempts to expose code, data and methods as tested tools an agent can invoke.

Does it work with every research paper?

Not automatically. It depends on accessible, runnable supporting material, compatible runtime requirements, useful examples and tests that verify the generated tools.

Why does MCP matter?

MCP standardises how compatible AI applications access resources, prompts and executable tools. The paper agent can therefore be connected to different clients rather than being locked to one chat interface.

Can a paper agent replace a domain expert?

No. It can reduce technical friction and surface alternative analyses, but experts remain responsible for evidence, assumptions and decisions.

UNDERSTAND
WHAT'S CHANGING.