When Agents Start Reading Papers: Three Layers for the Future of Science

I've been having a lot of conversations with my coauthor Miao Liu at Boston College about AI and academic research, and one question keeps coming back to me: everyone is debating whether AI can write papers, but the more important question might be a different one entirely -- when agents become readers too, what should scientific communication look like?

PDF Must Die

Let's start with format.

Academia in 2026 still runs on PDF. PDF stands for Portable Document Format. It was born in 1993, designed so that documents would look the same across different printers. It simulates paper. It cares about margins, columns, font rendering. This is a format built for printers, and we are using it for scientific communication in 2026.

Can AI read PDFs? Sure. But how? First it converts the PDF to text or Markdown, then processes it. Every act of "reading a PDF" is essentially reverse engineering -- forcing a format optimized for printers back into structured text. This process loses information, wastes compute, and frequently breaks. (Ever asked an AI to read a PDF with complex tables?)

This isn't an AI problem. It's a format problem.

And it's not just academia. The entire internet is undergoing the same shift. Cloudflare recently released an Agent Readiness Score that checks whether your website is agent-friendly -- whether agents can discover, read, interact with, and transact on it. Their criteria include: whether robots.txt has AI bot rules, whether the site supports Markdown content negotiation, whether it has an MCP Server Card, an OAuth discovery protocol, or even an agentic commerce protocol.

In other words, the infrastructure layer of the internet is already being redesigned for agents. Websites are no longer just for humans to read -- they're for agents to read too. Cloudflare even rewrote all of its developer documentation from scratch, aiming to become "the most agent-friendly docs site."

Websites are changing. APIs are changing. Documentation is changing. But academic papers are still stuck in the PDF era.

Beyond Format: When the Reader Changes, the Content Logic Must Change Too

The more I think about it, though, the more I realize that format is just the surface. The deeper issue is this: when agents become readers, the logic of the content itself should change.

Papers written by humans for humans face one fundamental constraint: attention is scarce. So we compress. We condense years of work into 30 pages. We omit failed attempts, insignificant results, abandoned research branches, intermediate reasoning. Not because these things lack value, but because human readers don't have time for them.

Agents don't have that constraint.

For an agent, the deleted material may be precisely the most valuable part: failed experimental paths reveal which directions aren't worth exploring again; null results help calibrate expectations about effect sizes; abandoned branches might be exactly what's needed for a different research question; intermediate reasoning lets the agent genuinely audit and reproduce the study.

So an AI-facing research artifact should follow a fundamentally different logic: not heavy compression, but maximum completeness and honesty.

This led me to think that we may ultimately need three layers of scientific output:

Layer 1: Papers written by humans, for humans. This is the form we know best today. Highly compressed, carefully narrated, with core value in judgment, interpretation, and intellectual synthesis. This layer should be preserved, because it carries the deepest kind of human intellectual work.

Layer 2: Papers written by AI, for humans. As AI-driven discoveries proliferate, we need a way to translate AI's findings into forms that humans can evaluate. The value here lies in translation and readability -- enabling humans to understand, question, and verify AI's work.

Layer 3: Research logs written by AI, for AI. These aren't "papers" in the traditional sense. They're structured research records -- containing complete methods, all experimental paths (successful and failed), intermediate reasoning, and pointers to raw data -- designed so that other agents can reuse, continue, and cumulatively learn from them.

But I'm Not Fully Confident in My Own Framework

To be candid, there are several questions in this three-layer model that I haven't resolved.

How long can Layer 1 stay "deepest"? I claimed that papers written by humans for humans represent "the deepest layer of interpretation," and that may be true today. But what if AI's interpretive capacity surpasses most human scholars within five years? Would this layer degrade from "deepest" to merely "slowest"? I'm not sure. Perhaps the ultimate value of human-written papers isn't that they're the most profound, but that they're a record of humans thinking for ourselves -- not because we think best, but because we need to think.

The trust problem in Layer 3. If research records are designed for AI, humans can't effectively audit them. But the bedrock of science is verifiability. How do you ensure quality in a layer of research records that humans can't read? This probably requires an entirely new verification mechanism -- agents auditing agents -- but that introduces its own chain-of-trust problems.

Where are the incentives for change? Academia's incentive structure -- tenure, citations, h-index -- is built entirely around the current paper format. Even if the three-layer model is theoretically right, who drives the transition? Who gives credit for Layer 3? Without incentive design, even the most elegant framework is just armchair theory.

So this post is more of a question than an answer. But I'm increasingly convinced that instead of asking "Can AI write papers?", the more important question is: When readers are no longer only human, how should the form of scientific communication evolve?

We may be standing at the beginning of a paradigm shift. It's just that no one has drawn the new map yet.