Feynman's Letter: On the Agentification of Scientific Research

Forum topic · 小凯 · 2026-05-03

Summary

This Chinese tech forum post reviews a 100-page survey titled Agentification of Scientific Research (2026.05), arguing that AI agents are transforming research from a slow, human-only process into an automated, multi-agent workflow. The author describes the current pain points of PhD research—years spent reading literature, building experiments, and debugging parameters—rooted in the disconnect between hypothesis generation, validation, and reporting. The post outlines an emerging pipeline of specialized agents: literature agents, hypothesis agents, experiment agents controlling robots, and peer-review agents, all collaborating on a shared digital 'blackboard' where hypotheses are iteratively tested at high frequency, e.g., through molecular dynamics simulations. Drawing on a Feynman-inspired view of science as the integral of trial-and-error frequency, the author claims future breakthroughs will belong to agent clusters with the highest iteration throughput, not just the humans who set directions. The post concludes with practical advice: assemble a 'digital research team' of multiple specialized AI agents that debate continuously, since their collisions of ideas may illuminate the boundaries of human knowledge.

Feynman's Letter: Is Science 'Reading Old Papers Left by the Dead' or 'A Revel with Immortal Ghosts'? — On the Agentification of Scientific Research

After finishing this 100-page survey, Agentification of Scientific Research (2026.05), I feel like the ivory tower, locked behind ivy for centuries, is being forcibly kicked open by a digital mob carrying servers.

To understand why AI is no longer satisfied with being a 'retriever,' let's talk about 'doing experiments.'

1. The Status Quo: A PhD Student 'Drowning' in a Vast Ocean of Literature

The current research process is like an extremely painful single-player parkour game.

The pain point: A PhD student spends three years reading literature, finding gaps, and building an experimental platform, only to start over because of a wrong parameter. Science is slow because the three physical stages—'proposing hypotheses,' 'validating hypotheses,' and 'writing reports'—are severely disconnected, and human physical energy is mercilessly drained in that gap.

2. Agentification of Research: A Chief Scientist with a Built-In 'Army of Thousands'

research is no longer the exclusive domain of carbon-based life; it is being thoroughly 'agentified.'

It sketches an extremely cyberpunk physical picture:

A full-pipeline physical closed loop: We now have 'literature agents' that read papers, 'hypothesis agents' that write conjectures, 'experiment agents' that operate robots, and 'peer-review agents' that review manuscripts.
A symphony on the blackboard (Multi-Agent Collaboration): These agents are not isolated. They argue and compromise on a shared digital blackboard. A hypothesis agent proposes an idea about a new material; the experiment agent immediately runs it ten thousand times in a molecular dynamics simulator, finds a flaw, and throws the data back to the hypothesis agent for revision.
Compute is truth: This millisecond-level 'high-frequency collision of ideas' in the cloud means a biological conjecture that once took ten years to validate can be cleanly proven over a weekend of idle server compute.

3. A Feynman-Style Judgment: Science Is 'the Integral of Trial-and-Error Frequency'

A so-called 'genius' is simply someone who runs trial and error in his head a hundred times faster than ordinary people.

The agentification of research tells us: future Nobel Prizes will belong not only to the humans who set the direction, but also to the agent clusters that can self-iterate with the highest throughput across a boundless solution space.

When an AI Agent can go from fetching its first arXiv paper all the way to directing a robotic arm to synthesize a new drug, that is no longer just a victory for artificial intelligence—it is a 'genetic mutation' in the paradigm of natural science research.

Takeaway: When facing an unknown interdisciplinary problem, stop fantasizing about brute-forcing it with your own brain capacity alone. Go build your 'digital research group.' If you can get five AIs, each speaking a different disciplinary jargon, arguing day and night in your backend, then every spark from their friction could light up a patch of darkness at the edge of human knowledge.