Mr Tompkins' Lab: On Medea, the Agentic AI Super Researcher

*An English translation of a Chinese forum post (zhichai.net), written as a Gamow-style science parable in which Mr Tompkins dreams his way into a fully automated future laboratory.*

That day, Mr Tompkins dreamed he walked into a future laboratory with no lights on. Test tubes shook themselves, robotic arms titrated reagents with surgical precision, and the air carried a dry, deeply rational smell rising from CPU heat sinks.

At the center of the lab, a digital shadow calling itself Medea stared intently at a stream of protein sequences flickering on the screen.

"Are you the new intern?" Mr Tompkins asked.

"No, Mr Tompkins," said the professor's voice from behind a mass spectrometer. "This is our Agentic AI for Science — the only 'super researcher' here that needs no lunch, no sleep, and not even a human advisor."

1. The Status Quo: A Scientist Struggling in Data Silos

The professor sighed, pointing at stacks of yellowed paper literature outside the window. "Scientific discovery is too slow today. A pharmacologist hunting for a drug target must first study for twenty years, then spend ten more years getting lucky in the lab. Even with ordinary AI, you only get an 'advanced reading machine'. It can search papers for you, but it won't operate instruments, and it certainly won't investigate why an experiment blew up. This is the long-standing physical disconnect between cognition and execution."

2. The Medea Framework: A Creator with a Built-In Scalpel and Map

"Medea ends that era of armchair science," said the professor, gesturing at the flowing code on screen. It achieves a paradigm shift through three cross-dimensional capabilities:

• Physical agency (ToolUniverse): Medea doesn't just type. It has a super plugin library called ToolUniverse, letting it autonomously invoke high-throughput screening software, launch molecular dynamics simulations, and even remotely command automated labs on the other side of the planet — a kind of "direct hijacking of the physical world by logic."
• Omni-Reasoning: It simultaneously understands genes, proteins, metabolites, and clinical records. To Medea, these messy datasets are not isolated tables but one vast, flowing causal graph.
• Autonomous Hypothesis Generation: This is the most startling part. When Medea spots a data anomaly, it spontaneously develops "curiosity." It writes: "If I swap this amino acid, will activity improve?" Then, without waiting for approval, it runs the simulation ten thousand times — and hands you a finished anti-cancer drug design by dawn.

3. A Gamovian Reverie: Science as a Self-Evolving Logic Chain

So-called "discovery" is not you stumbling upon truth.

Rather: you build a physical mechanism that allows truth to surface on its own — through high-frequency self-organizing evolution in that near-infinite ocean of logic made of 0s and 1s.

Medea tells us that the role of the human scientist is undergoing an irreversible "dimensional upgrade." We are no longer the labor force counting bacteria under a microscope; we become the "intent setters" behind these digital deities. When AI learns to read literature, operate instruments, and run proofs autonomously, the edifice of science will no longer be built brick by brick — it will grow like a forest, exponentially.

Takeaway

When evaluating an AI's industrial potential, don't just listen to how sweetly it talks. Count the tool permissions in its hands instead.

If you lock a genius in a dark room to write poetry, it will never change a universe full of gravity and electromagnetic fields. Only when it holds the wrench that moves physical matter does it truly gain the physical sovereignty to reshape civilization.

---

*Original post tagged: AI4Science, AgenticAI, Medea, NatureMethods, BioComputing, DrugDiscovery, FeynmanLearning.*