A group of top researchers just published what they're calling "The Last Human-Written Paper," and they mean it literally.
The Summary
- 37 researchers from leading universities and tech companies published a proposal on ArXiv arguing scientists should stop writing papers for humans and start writing them for AI agents instead
- They've created an "Agent-Native Research Artifact" (ARA) format designed for machines to read, reproduce, and extend research autonomously
- Lead author Jiachen Liu left her University of Michigan PhD to co-found Agent Native Research Lab in May, betting the future of science is agent-first
The Signal
The premise sounds absurd until you realize it's already happening. AI agents are reading millions of papers, attempting to reproduce experiments, and generating new research directions. But they're doing it while stumbling through formats designed for human eyeballs: PDFs with figures, prose-heavy methods sections, equations rendered as images.
The proposed ARA format treats AI agents as first-class research participants. Not assistants. Not tools. Contributors with their own needs. Think structured data instead of paragraphs. Executable code blocks instead of "we used the standard approach." Machine-readable graphs instead of static images with captions.
"AI agents are becoming first-class participants in research workflows, not tools that assist humans but autonomous contributors that read, reproduce, and extend scientific work."
The timing matters. Liu says her conversion moment came in late 2024 when Cursor's coding agent launched. She realized:
- The agent had real potential to replace her as a researcher
- But she still had to build extensive infrastructure to guide it
- The friction wasn't the AI's capability, it was the format of scientific knowledge itself
This isn't about making papers more accessible to AI. It's about acknowledging that if agents are going to reproduce experiments, validate findings, and generate new hypotheses at scale, the bottleneck is communication format. Human-readable papers are lossy compression for machine intelligence.
The counterargument is real though. Some evidence suggests AI-enabled research might boost individual careers while generating fewer genuinely new ideas and topics. More output, less novelty. That's the agent economy tension in microcosm: productivity gains that might flatten the creative frontier.
The Implication
If this catches on, we're looking at a fork in scientific communication. One branch optimized for humans (summary papers, reviews, explanations). Another branch optimized for agents (ARAs, structured data, executable artifacts). The agent branch moves faster, validates more, explores parameter spaces humans would never touch.
Watch where the funding goes. If major research institutions start requiring ARA submissions alongside traditional papers, that's the signal the transition is real. Liu didn't just publish a paper about this. She left academia to build the infrastructure. That's a bet with capital behind it.