Researchers Propose Agent-Native Artifacts as a New Format for Scientific Work

IEEE Spectrum has highlighted a 37-author proposal that challenges a basic scientific convention: researchers should stop treating human-readable papers as their primary output. The authors want machine-executable research packages built for AI agents instead. Their argument is not that people should lose access to science. It is that artificial intelligence now needs more detail than a conventional paper can carry.

The proposal appears in The Last Human-Written Paper, submitted to ArXiv on April 27, 2026, and revised on May 19. Its authors call their replacement an Agent-Native Research Artifact, or ARA. The format combines scientific claims, executable code, research history, and raw evidence inside one structured package.

That pitch creates a sharper conflict than the provocative title suggests. Science must become easier for machines to inspect without becoming harder for people to question. The real contest is therefore not AI readers against human readers. It is machine efficiency against the narrative, editorial, and social functions that make research understandable.

The paper reports substantial gains on selected machine-learning benchmarks. Yet its own results also reveal a warning. Detailed records of previous failures can guide an AI agent toward useful work, but those records can also anchor it to outdated ideas.

What the IEEE Spectrum Story Actually Put on the Table

The proposal changes the primary scientific object from a persuasive document into an operable package.

A conventional paper presents a research project as a linear argument. It introduces a problem, describes a method, reports results, and explains why those results matter. Code, data, supplementary files, and laboratory records usually sit outside that central narrative.

ARA reverses that hierarchy. The structured package becomes the main record, while a readable paper becomes one compiled view of its contents. An AI agent can inspect individual claims, trace evidence, execute code, or follow the decisions that shaped an experiment.

The 37 researchers organize ARA into four connected layers. A cognitive layer records the problem, concepts, experiments, and falsifiable claims. A physical layer contains executable code, configurations, dependencies, hardware requirements, and seeds.

An exploration graph preserves branches, rejected approaches, failures, and pivots. An evidence layer holds results, logs, curves, and other outputs supporting each claim. Cross-references connect an assertion to its experiment, code, and evidence.

This structure responds to two problems that the authors name the Storytelling Tax and Engineering Tax. The Storytelling Tax arises when researchers compress months of branching work into one successful path. Failed experiments and rejected hypotheses often disappear during that compression.

The Engineering Tax describes the difference between information sufficient for persuasion and information sufficient for execution. A reviewer might understand why a method should work without possessing every setting needed to reproduce it. An AI agent attempting the experiment cannot bridge that gap reliably.

The authors also propose three supporting systems. A Live Research Manager would capture decisions during ordinary work with an AI assistant. An ARA Compiler would convert existing papers and repositories into the new structure.

An ARA-native review process would then check structural integrity, argumentative rigor, and reproducibility. Human reviewers would remain responsible for judgments involving significance, novelty, and scientific taste.

The IEEE Spectrum analysis makes this proposal newsworthy because it moves beyond using AI to summarize papers. It asks whether scientific communication should be redesigned around autonomous systems that read, reproduce, and extend research.

That distinction matters. Search tools treat papers as documents to retrieve. ARA treats research as an environment in which an agent can act.

The proposal also goes further than attaching a notebook or repository to a PDF. Its central claim is that narrative papers are lossy compilations. Adding more supplementary files does not necessarily restore the connections between claims, decisions, implementation details, and raw evidence.

ARA therefore makes a direct institutional demand. Universities, conferences, journals, funders, and research teams would need to recognize structured artifacts as first-class outputs. Without that recognition, ARA remains an elaborate supplement rather than a replacement.

Why AI Agents Put Scientific Publishers Under Pressure

Publishers now face a format problem because AI systems consume research differently from human readers.

People usually begin with titles, abstracts, figures, and conclusions. They move between sections according to interest and expertise. A well-written narrative helps them understand why a question matters and how evidence changes an argument.

An AI agent performing technical work has different needs. It must identify precise claims, resolve dependencies, locate configurations, inspect code, and connect reported numbers to raw outputs. It also needs to distinguish a confirmed result from a hypothesis or abandoned idea.

A PDF can contain those details, but it rarely exposes them consistently. Tables may be difficult to parse. Equations can lose structure during extraction. A method may depend on settings scattered across prose, appendices, code, and undocumented conversations.

The ARA researchers examined 8,921 expert-annotated reproduction requirements associated with 23 ICML 2024 papers. They report that only 45.4 percent were fully specified in the corresponding PDFs. Code-development requirements had 37.3 percent sufficient coverage.

Missing hyperparameters accounted for 26.2 percent of the identified information gaps. These figures come from the proposal authors’ analysis, not an independent audit of ARA. Still, they describe a familiar reproducibility problem in unusually concrete terms.

The underlying benchmark is PaperBench, which evaluates whether AI agents can reproduce machine-learning research. Reproduction involves more than answering questions about a paper. The agent must rebuild experiments and produce outputs that expert graders can assess.

Publishers are not the only groups under pressure. Researchers would have to capture more structured information while a project is active. Laboratories would need policies for preserving internal decisions without disclosing confidential material.

Conference reviewers would need tools for navigating artifacts that may contain thousands of files and events. Archivists would need stable standards for environments, dependencies, provenance, and access permissions. Research institutions would also need to decide which machine actions count as valid verification.

The paper’s answer to the authoring burden is automation. Its Live Research Manager would observe human-agent work and turn decisions into structured records. That approach assumes an AI assistant participates throughout the project.

This assumption fits some computer-science laboratories. It fits wet-lab biology, clinical studies, field research, and qualitative work less naturally. Much of their evidence begins as physical observations, interviews, instruments, samples, or restricted records.

The authors acknowledge that their evaluation covers machine learning. They say the physical and exploration layers may require substantial adaptation for experiments conducted outside computational environments. That limitation narrows the immediate claim considerably.

Scientific publishing also serves readers who do not execute code. Policymakers, journalists, students, interdisciplinary researchers, and affected communities need explanations, not repository traversal. A format optimized only for agents would fail these audiences.

The stronger version of the proposal is therefore dual publication. A structured research object can hold executable detail, while a human narrative explains purpose, uncertainty, and consequence. The paper itself effectively supports this approach by calling narratives compiled views.

IEEE Spectrum is exposing a real pressure point, but the practical response does not require abandoning prose. It requires ending the PDF’s role as the only authoritative representation of a project.

IEEE Spectrum’s Core Conflict Is Knowledge Versus Narrative

Machine-readable structure can preserve more research detail, but detail alone does not produce scientific understanding.

The authors summarize their design principle as “Knowledge over Narrative.” That phrase identifies the proposal’s strongest insight and its central weakness. Research knowledge is not always separable from the way an argument is constructed.

Narrative can hide failures and smooth over uncertainty. It can also explain why a problem deserves attention, which assumptions carry moral weight, and where evidence remains ambiguous. Those functions are difficult to reduce to claims, dependencies, and executable tasks.

Scientific papers are persuasive because science requires judgment. Readers assess whether a question is important, whether a comparison is fair, and whether an explanation fits the evidence. A machine-verifiable package can support those decisions without making them automatic.

ARA’s four layers nevertheless offer clear benefits. Claim-to-evidence links can expose unsupported conclusions. Configuration records can reduce wasted reproduction attempts. Exploration histories can stop multiple teams from repeating the same failed approach.

The proposed PAPER.md entry point also supports progressive disclosure. An agent can first read a compact overview, then load only the files needed for a particular question. This design helps control context use while preserving deeper material.

The approach extends earlier efforts to make research objects easier to find and reuse. The FAIR principles established that data should be findable, accessible, interoperable, and reusable. Other systems package workflows, provenance, notebooks, or atomic claims.

ARA attempts to connect all those dimensions with execution and decision history. Its novelty lies less in any single file type than in the bindings between them. A claim can point to an experiment, code path, output, and sequence of prior decisions.

That structure would make scientific work easier to audit if researchers maintain it faithfully. It could also make selective reporting more visible. A reviewer could inspect how many approaches failed before the reported method succeeded.

However, exhaustive records introduce new interpretive problems. A failed experiment might reflect a flawed implementation rather than a bad hypothesis. An agent could treat an old failure as a universal constraint when it only applied to one environment.

The paper’s own extension tests demonstrate that risk. ARA-equipped agents found useful first moves sooner across five RE-Bench tasks. They had access to failure traces and heuristics that paper-based agents had to rediscover.

On two tasks using a more capable model, the early advantage did not produce the best final result. The paper-based agent kept exploring beyond the documented playbook. The ARA agent remained anchored to strategies recorded in the artifact.

That outcome weakens the simplest sales pitch. More structured context did not always produce better open-ended research. Sometimes ignorance preserved a wider search space.

The relevant tradeoff is between memory and exploration. Scientific systems need enough memory to avoid needless repetition. They also need mechanisms that prevent inherited conclusions from becoming invisible boundaries.

Human researchers already face this problem. Literature reviews help them build on previous work, but accepted frameworks can discourage alternative questions. A machine agent could amplify that effect because it follows explicit structures consistently.

ARA could address the risk by attaching provenance, dates, model capabilities, and scope conditions to failed branches. Agents could discount an old result when hardware, data, or methods have changed. They could also receive deliberate exploration budgets outside the recorded graph.

Those protections are not minor implementation details. They determine whether an agent-native artifact becomes a map or a cage. A map identifies known terrain while leaving room beyond its borders.

The best outcome would preserve both representations. Humans would receive a clear narrative with accountable authorship. Agents would receive an inspectable research object containing the operational detail that prose cannot efficiently hold.

The Benchmark Gains Are Significant but Not Yet a Verdict

The reported results justify further testing, not a general replacement of scientific papers.

The authors evaluate understanding, reproduction, extension, and automated review. Their most direct comparison gives agents either an ARA package or a conventional paper with associated code. The tests then measure how well each representation supports research tasks.

For question answering across PaperBench and RE-Bench, the paper reports accuracy rising from 72.4 percent to 93.7 percent. That is a 21.3 percentage-point difference across 450 questions. The result suggests structured artifacts help agents locate precise technical information.

For reproduction, the reported success rate rises from 57.4 percent to 64.4 percent. The evaluation covers 150 subtasks drawn from 15 PaperBench papers. The seven-point improvement is useful, but it still leaves more than one-third of tasks unsuccessful.

The extension results are more complicated. ARA achieved the best score on three of five open-ended tasks and found an effective initial direction sooner on all five. Yet the prior failure records sometimes constrained stronger agents.

These are authors’ results from a preprint. The work had not passed a traditional journal review when posted to ArXiv. Independent teams still need to reproduce the comparisons and test whether the artifact construction process favored the ARA condition.

The paper discloses several important limits. Evaluators familiar with the selected papers and ARA format helped build the human-annotated benchmark. Performance may differ for unfamiliar subjects, unusual methodologies, or less curated artifacts.

The compiler also cannot recover information absent from an original paper. If authors omitted an experiment setting, an extraction system cannot reconstruct it reliably. A converted legacy paper therefore inherits many of the source document’s gaps.

Live capture can preserve more information, but it creates another dependency. The system assumes research decisions pass through an AI-mediated workflow. Conversations outside that workflow, manual experiments, private messages, and unrecorded judgments can remain missing.

Security is a larger unresolved issue. The paper says production requirements including sandboxed execution, content-level anomaly detection, and granular access controls are not yet implemented. Those controls are necessary before agents execute unfamiliar research packages.

A malicious artifact could contain unsafe code or misleading instructions. A compromised dependency could alter results. Sensitive traces might expose unpublished ideas, personal information, security findings, or confidential datasets.

Privacy also conflicts with comprehensive capture. Preserving every rejected idea and conversation sounds valuable for machines. Researchers may speak less freely if each speculative comment becomes part of a durable artifact.

Selective access can reduce that concern, but restricted evidence creates verification complications. Reviewers must know what was withheld and why. Institutions must also prevent private material from leaking through generated summaries or derived outputs.

Automated review introduces another uncertainty. The proposed ARA Seal checks structure, rigor, and execution at increasing levels. Structural completeness, however, does not guarantee that a claim is scientifically important or conceptually sound.

Large language models can also make inconsistent judgments. The paper reports pathologies involving its model-based rigor auditor, including mismatches between detected issues and assigned scores. Human review remains necessary even inside the proposed architecture.

These limitations do not erase the gains. They show why adoption should begin with parallel artifacts in computational fields. Conferences can request ARA packages while retaining human-readable papers and normal review.

Teams could then compare authoring burden, reviewer time, reproduction rates, security incidents, and artifact reuse. Those measurements would provide stronger evidence than replacing papers based on benchmark performance alone.

The proposal is best understood as infrastructure research. It offers a testable format and supporting tools. It has not established a universal endpoint for scientific communication.

Three Signals Will Show Whether Agent-Native Research Works

The next phase should be judged by adoption, independent reproduction, and operational safety rather than provocative titles.

The first signal is adoption by conferences, journals, or major laboratories. A credible pilot would require structured artifacts for a defined class of computational submissions. It would also specify archival rules and reviewer responsibilities.

Voluntary GitHub uploads are useful, but they do not test institutional acceptance. Researchers respond to publication requirements, citation practices, funding rules, and promotion incentives. A format becomes infrastructure only when those systems recognize it.

The second signal is independent reproduction. External teams should rebuild the reported ARA comparisons using unseen papers, different agents, and multiple model families. They should disclose the cost of creating and maintaining each artifact.

A useful study would separate three effects. One comes from better formatting. Another comes from adding information unavailable in the paper. The third comes from evaluators knowing how the artifact was constructed.

That separation matters because any detailed research package might outperform a short PDF. The harder question is whether ARA’s specific ontology provides durable benefits across projects and institutions. Independent tests can establish that difference.

The third signal is a credible security and governance model. Before agents execute shared artifacts, the ecosystem needs sandboxing, dependency controls, provenance checks, permissions, and durable audit logs. These safeguards must work across organizations.

Governance must also define authorship. If agents capture decisions, modify claims, execute experiments, and generate summaries, readers need to know which contributions came from people. Attribution cannot become an opaque field generated after the work is complete.

The public ARA repository provides code and examples for experimentation. That openness can help independent researchers inspect the proposed schema. It also gives potential adopters a concrete object to test rather than a purely theoretical standard.

Near-term adoption will probably be uneven. Machine learning and computational systems offer the clearest fit because their experiments already involve code, configurations, logs, and measurable outputs. Other disciplines will require different physical and ethical controls.

Researchers should not interpret the IEEE Spectrum question as a command to write for machines instead of people. They should treat it as a demand to stop forcing one document to serve every audience.

A paper can remain the human interface. An agent-native package can become the operational record. Neither should silently replace the other until evidence shows which scientific functions each format performs well.

This division also offers a practical workflow for knowledge workers. Teams can preserve structured source material while producing readable explanations for colleagues. A searchable knowledge base follows the same broad principle: retain the evidence while making retrieval easier.

The difficult part is not generating more files. It is maintaining trustworthy connections among claims, decisions, code, and evidence. Those connections must survive software changes, personnel turnover, and disagreements about interpretation.

IEEE Spectrum has surfaced a serious proposal beneath an intentionally dramatic title. The future of scientific publishing is unlikely to be the last human-written paper. It is more likely to be the first research record designed honestly for both kinds of reader.

Researchers should ask a concrete question before adopting ARA: does this package make the work easier to verify without weakening human understanding? If independent replications, institutional pilots, and secure execution answer yes, agent-native artifacts deserve a formal place beside papers. Until then, scientists should experiment with the format, preserve readable narratives, and resist treating machine efficiency as the only measure of scientific communication.