Steven Watson | EdgeLab | Research and Scholarship | 19 September 2026

What would change if reading a research paper also meant being able to ask its methods to do something? The attraction is clear: less time negotiating unfamiliar software and more opportunity to explore a question. But when a document becomes an executable service, access, judgement and responsibility have to be considered together.

What Paper2Agent demonstrates

In a peer-reviewed Nature article published on 16 September 2026, Jiacheng Miao and colleagues introduce Paper2Agent. It packages research code and resources as tools that conversational AI can invoke through the Model Context Protocol, a shared interface for connecting models to external functions. Automated testing checks outputs against reference examples.

Across 100 computational-biology papers, 74 were successfully converted into executable agents. On 300 tutorial-derived questions from that successful subset, reported accuracy was 91.2%, versus 80.3% for the same underlying model with direct repository access. Further tests covered other computational fields and information synthesis.

The authors explicitly retain human responsibility for open-ended interpretation. Their discussion distinguishes faithful execution from analytical validity and acknowledges continuing maintenance. This matters: the study is not evidence that an agent can authoritatively settle any scientific question.

Two different meanings of getting it right

My reading separates an engineering achievement from a judgement about knowledge. Reproducing an expected result is valuable. Deciding whether that result warrants a particular conclusion is another task. A researcher can execute a method flawlessly and still apply it to an unsuitable question. Conversely, a defensible reanalysis may disagree with a previously reported interpretation.

This distinction suggests two complementary evaluations. First, can a system carry out a specified procedure accurately and make its actions inspectable? Second, can users recognise when the procedure is inappropriate, when a choice needs justification and when independent expertise is required? I would not want success on the first question to stand in for an answer to the second.

The denominator deserves equal attention. A service’s reliability for supported tasks and the proportion of research it can support are separate properties. In adoption decisions, both matter. An institution should ask which materials remain difficult to reuse and whether the exclusions systematically disadvantage particular methods, fields or research communities.

A plausible alternative reading is that these are ordinary software-engineering concerns, not a fundamentally new epistemic problem. That objection has force. Better interfaces need not change the meaning of scientific justification. What makes the issue worth investigating is whether conversational access changes how people notice assumptions, challenge defaults and distribute responsibility. Those are empirical questions about use, not consequences we can read off the architecture alone.

An AE interpretation: what changes at each handoff?

The following interpretation draws on autopoietic ecology, rather than attributing AE commitments to the authors. A research result, a software function and a conversational answer do not have identical standing. Each becomes usable within a different arrangement of people, records, permissions and expectations.

EdgeLab’s research programme calls attention to semantic transduction: what changes when an expression moves between settings. Here, a useful inquiry would follow a user’s question into a selected analysis and then into a written conclusion. Which qualifications survive? Which defaults become invisible? Who can reopen a decision?

Consider a hypothetical researcher who requests a comparison across two datasets. The system completes it and supplies an attractive figure. The scientific issue is not only whether the plotting code ran. It is whether the comparison is meaningful, whether exclusions were justified and whether the resulting claim carries more certainty than the evidence permits. This example is invented; it is not an observed failure in the study.

The Concept Guide distinguishes activity being possible now from renewal of its supporting capacities. Making an analysis easy today is one achievement. Sustaining the documentation, infrastructure and expertise that make it trustworthy next year is another. Neither automatically entails the other.

This is also relevant to EdgeLab’s proposed AEAI work on constraint continuity: whether a restriction keeps its practical effect as tasks move between interpretation, delegation and execution. For example, permission to inspect a dataset should not silently become permission to upload it elsewhere. That is a proposed evaluation criterion, not a claim that Paper2Agent violates it or that AE supplies a finished solution.

What I would test next

A useful deployment study would involve researchers with different levels of domain and programming expertise. Give them tasks containing both straightforward analyses and cases where the right response is to question the request. Measure not only completion but whether they can explain their choices, identify uncertainty and recover from a mistaken assumption.

I would also test interruption and correction. If a user withdraws permission halfway through, what stops? If an upstream source is corrected, which derived outputs need review? If a tool becomes unavailable, is the limitation made clear or hidden behind a plausible answer? These tests would examine accountable use alongside task performance.

The institutional question is who has the resources to maintain and contest the resulting service. A nominal human supervisor is not enough if they lack time, access or the ability to change what happens. Oversight should be demonstrated in action, including occasions when people refuse a proposed analysis.

The constructive promise is easier, more inspectable reuse of research. The condition is that convenience should strengthen rather than obscure scientific judgement. We should welcome tools that expand what researchers can do while continuing to ask how people can understand, challenge and take responsibility for what is done.

Reference

Miao, J., Davis, J. R., Zhang, Y., Pritchard, J. K., and Zou, J. (2026). Reimagining research papers as interactive and reliable AI agents. Nature. Published 16 September. doi:10.1038/s41586-026-11044-y. The publisher identifies reviewers and provides a peer-review file. This review is based on the full article and methods, not the accompanying news coverage.