Science Is Entering Its Agentic Era

Science has always been bottlenecked by human time. PubMed holds more than 38 million papers. ClinicalTrials.gov registers more than 500,000 trials. The combinatorial space of chemistry and biology exceeds what any individual researcher, or any team, can navigate in a lifetime. This has been the quiet ceiling on scientific velocity for decades.

Large language models cracked a first layer of that ceiling in 2023. Tools like Consensus, Elicit, and Scite turned static literature into conversational knowledge. Useful, but incremental: they summarized what was already known.

What is happening now is different in kind. We are moving from assistive AI that helps humans read faster, to agentic AI that conducts research autonomously: formulating hypotheses, designing experiments, executing code, reading full-text papers through tools, and writing manuscripts. The loop is starting to close.

Signals From the Last Twelve Months

A few facts worth noting:

FutureHouse, the AI-for-science lab co-founded by Sam Rodriques and Andrew White, launched its public platform on May 1, 2025, with five specialized agents (Crow, Falcon, Owl, Phoenix, and Finch), each built for a distinct stage of the research process. A commercial spinoff, Edison Scientific, followed in November 2025.
In May 2025, FutureHouse demonstrated a multi-agent workflow that identified a novel therapeutic candidate for dry age-related macular degeneration, a leading cause of irreversible blindness.
Sakana AI's AI Scientist autonomously generates research ideas, writes code, runs experiments via agentic tree search, produces full manuscripts, and performs its own peer review. In its template-free mode, the system submitted three papers to an ICLR 2025 workshop. One would have been accepted had it not been withdrawn per protocol. The paper's scientific quality improves predictably with both compute budget and base model capability: the exact scaling dynamic that turns early demos into massively-adopted tools.
Argonne, Oak Ridge, and Lawrence Berkeley National Laboratories now operate autonomous, AI-driven materials laboratories.

None of this is finished work. These agents still hallucinate citations and miss methodological subtleties. But machine learning follows a consistent pattern: once a capability starts to work (even poorly), scale and better base models tend to push it past human performance faster than most people expect.

What It Actually Means

The right analogy is not ChatGPT. It is the printing press, or the moment computational simulation entered the physicist's toolkit.

Three compounding effects are plausible over the next five years:

Throughput. Literature review and hypothesis generation collapse from weeks to minutes, letting individual researchers explore ten to fifty times more paths in the same calendar time.
Coverage. Sub-fields too small or too niche to attract sustained human attention become economically searchable.
Integration. Dry-lab and wet-lab loops begin to close, with agents triggering real experiments, reading real results, and updating their own models.

The AI Scientist makes these dynamics tangible. A single run (about fifteen hours and twenty dollars of compute) covers ideation, implementation, experimentation, and write-up. Its template-free mode surfaces research directions a human researcher might never prioritize. And its agentic tree search closes the loop entirely in silico: generating code, executing it, debugging failures, plotting results, and iterating — no human in the loop until the manuscript exists.

By 2030, AI agents will run fully autonomous loops of scientific knowledge production, from hypothesis to manuscript, with human or robotic intervention only where real-world experiments require it.

The risks are real: AI slop, reproducibility crises, misuse. But the counterfactual is not a pristine slow science; it is the current system, which already produces an overwhelming and partly unreliable literature. Agents that verify, cross-check, and stress-test other agents' output will matter as much as agents that generate.

Why I Am Moving in This Direction

My background is biosensing, wearable technology, and applied AI. For the past two months I have been building LocusLab, regulatory verification tooling for BioTech submissions. Useful work, but downstream of the actual bottleneck.

The bottleneck is the rate at which humanity produces validated knowledge. Agents that generate hypotheses, run experiments, and verify scientific claims act directly on that rate. Nothing else I could build moves the needle as much.

So that is what I am building.

Over the coming weeks:

Talking to active researchers (PhDs, postdocs, PIs) about where their workflows break with and without AI.
Contributing to the open systems shipping in this space, starting with FutureHouse and Sakana AI.
Shipping small, opinionated agents in domains I already know: biosensing, regulated biology, and scientific verification.

The first is already live: an open-source scientific claim verifier that beats naive Claude by 30 F1 points on SciFact (read the project write-up). More coming.

If you are working on something adjacent, I want to hear from you.

A Note On The Process

Whenever I catch weak signals that I may be heading in the wrong direction, I have learned (the hard way) not to push through them. I take time to write, to discuss with friends, and to brainstorm with Claude. Putting thoughts on a page forces clarity. Talking to people who know me breaks the echo chamber. Working with an AI that will push back honestly catches the rationalizations I can't catch alone.

Things worth doing always take longer than expected. Part of that time is execution; part of it is making sure you are executing on the right thing.

References

Lu, C., Lu, C., Lange, R. T., Yamada, Y., Hu, S., Foerster, J., Ha, D. & Clune, J. Towards end-to-end automation of AI research. Nature 651, 914–919 (2026). doi:10.1038/s41586-026-10265-5
Sakana AI. The AI Scientist — open-source code. github.com/SakanaAI/AI-Scientist
FutureHouse. Launching FutureHouse Platform: Superintelligent AI Agents for Scientific Discovery (May 2025).
MIT News. Accelerating scientific discovery with AI (June 2025).
Singularity Hub. What the Rise of AI Scientists May Mean for Human Research (February 2026).