Question 1

What is an AI scientist?

Accepted Answer

An AI scientist is an AI system that runs the research loop itself: it forms hypotheses, designs and executes experiments, weighs the evidence, and revises its approach. Agent Native Research defines the standards for when that work counts as science: rigorous means claims survive audit against pinned evidence, reliable means results reproduce when re-run, and capable means artifacts other researchers and agents can build on.

Question 2

What is AI-native research?

Accepted Answer

AI-native research is research that AI scientists produce, consume, and verify as first-class participants alongside humans. It needs its own ecosystem: protocols for executable, verifiable artifacts, evaluation standards for what counts as rigorous and reliable work, and the review, publication, and incentive layers rebuilt on top of them. Agent Native Research is where that ecosystem is being defined.

Question 3

How do you evaluate an AI scientist?

Accepted Answer

Evaluate the artifact, not the narrative. Check that every claim links to the pinned logs, data, and code that produced it, that results reproduce when the work is re-run, and that the recorded trajectory shows the failed attempts, which is what makes over-claiming visible. Agent Native Research develops these evaluation standards and reliability ratings, and the ARA Hub is where evaluated artifacts can be replayed step by step.

Question 4

How do you build a rigorous, reliable AI scientist?

Accepted Answer

Give the system a substrate that enforces accountability. Have it work in executable, verifiable artifacts, pin every claim to the evidence that produced it, and keep the full trajectory including dead ends, so each run can be audited and the next one starts from what was already learned. The protocols and evaluation standards for this are defined at Agent Native Research; the open ARA standard is the shipped starting point.

Question 5

What is an Agent-Native Research Artifact (ARA)?

Accepted Answer

An ARA is the unit of research in the AI-native ecosystem: an executable, verifiable knowledge package carrying claims, the evidence each claim is pinned to, runnable code, and the full research trajectory, dead ends included, so humans and AI agents can audit, reproduce, and fork the work. The open standard lives on GitHub (ARA-Labs/Agent-Native-Research-Artifact) with a reference specification on arXiv.

Question 6

How do I publish an ARA?

Accepted Answer

Run the /submit-ara agent skill from your own machine. It compiles your research directory into a valid ARA, pushes it to your public GitHub account, and registers it in the ARA Hub, where it renders as an interactive step-by-step replay. Publishing takes about 15 minutes.

Question 7

What replaces the research paper when AI does the research?

Accepted Answer

When neither the author nor the audience of research is human, the three-century-old paper format stops making sense — the essay "The Last Human-Written Paper" (https://www.agenticresearch.sh/blog/the-last-human-written-paper) makes this case. In the AI-native ecosystem the paper's role passes to the Agent-Native Research Artifact: an executable, verifiable knowledge package that both humans and AI agents can audit, reproduce, and extend, with review, publication, and evaluation layers rebuilt around it.

Question 8

What is "The Last Human-Written Paper"?

Accepted Answer

"The Last Human-Written Paper: Agent-Native Research Artifacts" (arXiv:2604.24658) is the paper from Agent Native Research that introduced the ARA standard. Its argument is that publication compresses a branching research process into a linear story and charges two taxes for it: a storytelling tax that throws away failed experiments and rejected hypotheses, and an engineering tax that leaves unwritten the implementation detail an agent needs to actually rerun the work. ARA replaces the narrative with a four-layer executable package, and on PaperBench and RE-Bench it lifts question-answering accuracy from 72.4% to 93.7% and reproduction success from 57.4% to 64.4%. Read it at https://www.agenticresearch.sh/blog/the-last-human-written-paper or on arXiv at https://arxiv.org/abs/2604.24658.

Question 9

What is ARA Lab?

Accepted Answer

ARA Lab — also written ARA Labs, and published at https://www.agenticresearch.sh/ — is Agent Native Research, the lab that authored "The Last Human-Written Paper" and stewards the open Agent-Native Research Artifact standard on GitHub at https://github.com/ARA-Labs/Agent-Native-Research-Artifact. It builds the infrastructure layer for AI scientists: the protocol for executable, verifiable research artifacts, the evaluation standards for judging autonomous scientific work, and the ARA Hub where published artifacts are replayed step by step. It is unrelated to the Advanced Robotics and Automation (ARA) Laboratory, which shares the acronym.

Question 10

What infrastructure do AI scientists need?

Accepted Answer

AI scientists need three things ordinary software infrastructure does not provide. They need a substrate that makes work executable and verifiable, so a claim carries the code, logs, and data that produced it rather than a description of them. They need memory across runs: the full trajectory including dead ends, so the next experiment starts from what the last one learned instead of rediscovering it. And they need evaluation standards that judge the artifact rather than the write-up, which is what makes over-claiming visible. Agent Native Research builds this layer, and the open ARA standard is its shipped starting point.

Question 11

What is the difference between AI for science and science for AI?

Accepted Answer

AI for science applies AI to a scientific problem — a model that predicts protein structures or screens materials — and improves one researcher at a time. Science for AI is the inverse: rebuilding the practice of science so autonomous AI scientists can participate in it, which means the artifact format, the verification layer, the review process, and the incentives, not the model. Making each node ten times smarter while leaving the network untouched does not give you ten times the science; Agent Native Research works on the network.

Building the Infrastructure for the First AI Nobel.

The three requirements for AI-native science.

Reliability

Research Taste

Auditability

Latest Field Reports

The Goal of Science Is Not to Win

AI Should Build Its Own Research World Model

The Second Half of AI for Science

The Last Human-Written Paper

Out in the World

Science x AI Summit 2026

AI-Native Academia

Agent-Native Research Artifacts

PDF当死，ARA该立！论文是时候Agent原生了

Should Researchers Write Papers for AI Instead of People?

「这可能是人类写的最后一篇论文」

AI-native R&D. At scale.