What Is AI-Driven Research? How Long-Running AI Agents Are Changing the Scientific Frontier (2026)

This is Hamamoto from TIMEWELL.

AI-driven research refers to a way of doing research in which AI agents take on the core of the work — generating hypotheses, reviewing literature, analyzing data, planning the next experiment — while humans set the questions, verify the results, and make the final decisions. Over the past year, AI has shifted from "a tool that answers instantly when you ask" to "a partner that settles in and investigates on its own for tens of minutes to several hours." In this article, I lay out that shift — from Deep Research-style autonomous investigation, to long-running agents tackling unsolved math problems, to self-driving labs that run experiments with almost no human involvement — using only primary sources such as the developers' official announcements. Along the way, I look at how the work of a human researcher changes, and how high schools can let students try this safely, writing from the standpoint of a company that builds AI agents in-house.

At TIMEWELL, our core business is building an AI agent that autonomously works through export-control classification. That is exactly why we feel, in our daily development, the gap between the "clever AI" you see in a demo and the "trustworthy AI" you can rely on every day at work. The changes happening in the world of research come into sharper focus when you view them with that gap in mind.

Let me start with three key points.

• AI-driven research is not about asking AI a one-off question. It is about letting an AI agent run the loop itself — searching, reasoning, and deciding the next move — while humans hold on to the questions and the verification.
• In 2025, agents arrived one after another: some, like Deep Research, investigate autonomously for tens of minutes; others spend hours to days tackling hard math problems.
• The same tools work in high school inquiry-based learning. But without a design that prevents "outsourcing the question" and "leaving sources unchecked," the learning hollows out.

What AI-driven research is, and how it differs from "using AI"

Let me start by sorting out the terms. Using AI in research is nothing new. Running statistics software, typing keywords into a paper search — these have been routine for a long time. What makes AI-driven research decisively different is that the loop of judgment at the heart of research — "what to ask," "which literature to read," "what the data implies," "what to try next" — is run autonomously by an AI agent. Rather than staying glued to every step, humans concentrate on direction, verification, and the final call.

For the broader field there is the term "AI for Science." That is an umbrella name for the large current of bringing AI into every scene of scientific research, and I cover the whole picture in a separate hub article on what AI for Science is. Think of the AI-driven research I dig into here as a narrower cut within that: "agents that autonomously run the loop from hypothesis to verification."

Laid out as a table, the difference looks like this.

Aspect	Conventional use of AI	AI-driven research
Role of AI	A tool that executes one human-directed task	An agent that advances search and analysis in multiple steps on its own
Runtime	Instant answers in seconds to minutes	Tens of minutes to hours, sometimes days
Human involvement	Operating at every step	Concentrating on setting questions and verifying results
Output	Fragments of an answer	Cited reports or plans for the next experiment

In Japan this current has begun to move as national policy. The Ministry of Education (MEXT) is running the "Program for Scientific Research Innovation through AI for Science" (nicknamed SPReAD), which plans to select on the order of 1,000 projects across two open calls, with direct-cost funding capped at 5 million yen each¹. The first call runs from April 17 to May 18, 2026, and the second from June 2 to July 3, according to the official schedule¹. It is designed to nudge researchers in every field to "first try using AI," a sign from the government that AI-driven research is not just a matter for a few large enterprises.

Investigating for tens of minutes on its own: the evolution of Deep Research

The easiest entry point to AI-driven research is the family of Deep Research features. OpenAI released Deep Research on February 2, 2025². The usage is simple: you just hand it a topic to investigate. The AI then follows hundreds of online sources on its own, analyzes and synthesizes their content, and produces a report of the kind a research analyst might compile. A single task takes roughly 5 to 30 minutes, during which a human can go do other work and wait².

How demanding a question can it handle? As a difficulty benchmark there is "Humanity's Last Exam," a set of over 3,000 hard, cross-disciplinary problems. Deep Research scored 26.6% here, the highest level at the time of release². There is the caveat that it has an edge over other models because it can use web search, but the number shows that tens of minutes of autonomous investigation reaches beyond mere summarization into hard problems.

What I value more than the score itself is the design philosophy of "returning answers with sources." Everyone has had the experience of an AI stating something with full confidence that turned out to be wrong. An investigation agent making explicit which pages it based its work on exists precisely to keep humans able to verify those errors. Put the other way around, the moment you skip the step of opening the sources to check them, Deep Research turns into nothing more than a persuasive prose machine. This line matters decisively in the high school inquiry-based learning I will touch on later.

From "answering in seconds" to "grinding for days"

Another major change is how long AI spends facing a single problem. The AI of the past prized instant answers. Now it is different. Agents have appeared that settle in for hours to days, investigating and solving.

The frontier of mathematics is a clear example. On May 14, 2025, Google DeepMind announced "AlphaEvolve," an algorithm-discovery agent that combines Gemini with evolutionary search³. Applied to more than 50 unsolved or hard math problems, it rediscovered the best known solution in about 75% of cases and improved on the previous best in 20%³. As concrete examples, it raised the lower bound of the kissing number in 11 dimensions (a geometry problem — how many mutually touching spheres you can arrange — studied for over 300 years) from 592 to 593, and for multiplying two 4×4 complex matrices it found an algorithm that beats the method considered best since 1969³. These are actual cases of AI hitting on a solution humans had not thought of.

The power of long-running operation also showed at the International Mathematical Olympiad (IMO). DeepMind's Gemini Deep Think scored 35 out of 42 at the 2025 IMO, fully solving 5 of 6 problems to reach the gold-medal standard. That result was officially graded and certified by the IMO organizers⁴. What is more, it wrote its proofs in natural language, just as humans do, rather than a formal mathematical language, and within the same 4.5-hour limit⁴. The year before, in 2024, AlphaProof and AlphaGeometry 2 reached silver-medal equivalent (28 points, 4 problems), but at the time the problems had to be translated into a formal language by hand, and the computation took 2 to 3 days⁴. In a single year it moved from "translate by hand, take days" to "in natural language, on the spot."

I should add that on the same 2025 IMO problems, an experimental reasoning model from OpenAI was also announced as gold-medal level, at 35 points with 5 of 6 solved. But this was self-assessed rather than officially graded by the IMO organizers, which is a different footing from DeepMind's official certification⁵. The person in charge described the model as one that "thinks for hours"⁵. Even for the same "gold-medal level," telling apart who did the grading is the trick to reading AI achievements correctly.

Lining these up makes it clear that the use of AI has widened from "a model that answers instantly in seconds to minutes" to "an agent that settles in on the scale of hours to days to investigate and reason." A division of labor is emerging: Deep Research-style tools suit the groundwork of literature review and hypothesis generation, while these long-running models suit exploratory solutions to the hardest problems.

Experiments that run without human hands: self-driving labs and lab automation

Up to here the story stayed on screen. The other frontier of AI-driven research is the self-driving lab, which automates the real experiments themselves. It refers to a combination of robotics and machine learning that repeats — proposing hypotheses, planning experiments, running them with robots, analyzing the data, and deciding the next conditions to try — with almost no human hands involved.

A frequently cited case is the "mobile robotic chemist" reported by the team of Professor Andy Cooper at the University of Liverpool. This robot ran autonomously for 8 days, around the clock, carried out 688 experiments, and found a photocatalyst mixture roughly 6 times more active than the initial formulation⁶. The original is a 2020 Nature paper, so it is a somewhat earlier result, but it vividly showed a picture of AI deciding "let's try this condition next" on its own while a robot works day and night⁶. Through 2025 and into 2026, next-generation self-driving labs — with modular hardware and Bayesian optimization or large language models built in — are being reviewed extensively in academic journals⁷.

On the hypothesis-and-design side, too, AI is beginning to take over the front end of experiments. AlphaFold 3, which predicts protein structures, was announced by Google DeepMind and Isomorphic Labs on May 8, 2024, and can predict not only proteins but their interactions with DNA, RNA, and small molecules. Officially, it is described as delivering at least a 50% improvement over existing methods in predicting molecular interactions⁸. In the life sciences, Evo 2, a genomic foundation model whose preprint Arc Institute and NVIDIA released in February 2025, learned from over 128,000 genomes and over 9.3 trillion bases and reportedly judged whether mutations in the breast-cancer gene BRCA1 are benign or pathogenic with over 90% accuracy⁹. Structure determination and mutation assessment that once took years are being replaced by AI predictions that are fast and highly accurate.

The clearest proof that this current is real may be the 2024 Nobel Prize in Chemistry. On October 9, 2024, half the prize went to David Baker for pioneering computational protein design, and the other half to Demis Hassabis and John Jumper for protein structure prediction¹⁰. It was the moment the highest honor in science acknowledged that AI is driving the scientific frontier.

How the researcher's role changes

Reading this far, it might look as though researchers' work is being replaced by AI. I do not see it that way. It is more accurate to see the center of gravity shifting rather than the role disappearing.

Speaking honestly from the standpoint of a company that builds AI agents in-house, long-running autonomous AI has its own weaknesses. It mixes in plausible falsehoods — hallucinations. It can run for tens of minutes on a slightly off premise and produce a precise but off-target conclusion. And it is hard for humans to trace afterward why it arrived at a given answer. When we build our export-control classification agent, we design it so that the final judgment always rests with the client's human export-control officer, precisely because we know these weaknesses in our bones. The faster something moves autonomously, the more you need a design that keeps verification and accountability on the human side.

If so, then the work that remains with researchers — indeed grows in weight — is the following: deciding what to ask; verifying the AI's conclusions against sources; drawing the lines of ethics and safety; and taking responsibility for the results. Even when AlphaEvolve updated a math record, it was humans who "posed the question, had the AI solve it, confirmed its correctness, and gave it meaning." Even when a self-driving lab ran 688 experiments, it is the researcher who decides what counts as a discovery. The more AI takes on the muscle of investigation and computation, the more the value rises of the backbone — the ability to pose questions and the ability to verify. That is how I read it.

This change is not confined to the world of research. In the corporate field too, a division of labor has already begun in which AI agents take on investigation and first drafts while humans handle direction and verification. What we drum into clients in our hands-on support is the importance of deciding in advance the "range you delegate to AI" and the "range humans keep hold of." In research or in practice, if you leave that ambiguous and let an autonomous AI run, you are left with a pile of unverifiable output in exchange for speed.

Trying AI-driven research safely in high school inquiry-based learning

Finally, a word about the classroom. Many of the tools introduced here are within reach of high school students. With Deep Research-style features, AI can do cited background research on an inquiry topic in tens of minutes. Used well, there is no more powerful partner for inquiry. But used wrongly, inquiry hollows out. I write in detail about combining generative AI with inquiry in the practical article on high school inquiry-based learning and generative AI, but let me narrow the essentials of letting students experience AI-driven research safely to three.

First, leave the question-setting to the students. The moment you outsource "think of an interesting topic" to AI, you hand over the tastiest part of inquiry. Putting into words what you are curious about, and why you want to know it, is work you must not have AI stand in for. This is a miniature of the "work that stays with humans" from the previous section. Keep what you delegate to AI to searching and organizing the question you have set.

Second, make students always open and check the sources. As I wrote earlier, Deep Research returns answers with sources so that humans can verify them. I want high school students to build the habit of, rather than accepting the AI's compiled text as is, actually opening the pages cited as evidence and cross-checking what they say. That extra step is what builds the ability to read information critically — the very ability most in demand in the era ahead.

Third, do not assess by the final product alone. A polished report may come out, but its appearance alone cannot tell you whether it is a wholesale copy from AI or the result of a student posing and testing questions. Have students record which parts they delegated to AI, which they verified themselves, and how their thinking changed, and include that process in the assessment. Designed this way, you can let students use frontier tools openly while preventing outsourcing.

Grounded in this thinking, at TIMEWELL we help design inquiry-based learning and support students through our WARP program for schools and educational institutions. Having been involved in developing more than 500 people, what we feel is that students lean in remarkably when they move from being used by AI to being the ones who make use of AI. On the national direction for high school reform, I have written up an explainer on the N-E.X.T. High School Vision, and on what "abilities AI cannot replace" actually means, an article on how high school AI education will change, so reading them together should help you grasp the full picture.

Summary

• AI-driven research is a style in which an AI agent autonomously runs the loop from hypothesis to verification, while humans keep hold of question-setting, verification, and the final decision.
• Deep Research follows hundreds of sources in tens of minutes and returns cited reports. AlphaEvolve updated math records, and at the IMO, AI reached the gold-medal standard (mind the difference between official certification and self-assessment).
• Self-driving labs run the cycle from hypothesis to experiment with almost no human hands. AI that takes over the front end of experiments, such as AlphaFold 3 and Evo 2, has entered practical use.
• The role does not vanish; its center of gravity shifts. The more AI handles investigation and computation, the more the value of posing questions and verifying rises.
• The same tools work in high school inquiry. Leave the question to the students, always have them check sources, and assess the process — and you can let them experience the frontier while preventing outsourcing.

The change underway in research is not a matter for distant specialists alone. AI that investigates in tens of minutes, and AI that grinds for days, are now within anyone's reach. That is exactly why what is asked of us is a deeply human judgment: what to have that power investigate, and where to check for ourselves. Next time you use it, try opening the sources beneath the AI's answer once. From there, the manner of doing research alongside AI should come into view.

---

References

Related articles

• What Is AI for Science? Explaining the Frontier Where AI Accelerates Discovery
• What Is the N-E.X.T. High School Vision? A Guide to MEXT's Grand Design for High School Reform
• What Are "Abilities AI Cannot Replace"? How High School AI Education Will Change
• High School Inquiry-Based Learning × Generative AI: Practical Steps to Empower Students

Footnotes

1. MEXT, "Program for Scientific Research Innovation through AI for Science (SPReAD)" ↩ ↩²

2. OpenAI, "Introducing deep research" (February 2, 2025) ↩ ↩² ↩³

3. Google DeepMind, "AlphaEvolve: a Gemini-powered coding agent for designing advanced algorithms" (May 14, 2025) ↩ ↩² ↩³

4. Google DeepMind, "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad" (July 21, 2025) ↩ ↩² ↩³

5. Noam Brown (OpenAI), post on the 2025 IMO result (July 19, 2025) ↩ ↩²

6. B. Burger et al., "A mobile robotic chemist," Nature 583, 237–241 (2020) ↩ ↩²

7. Chemical Reviews, "Self-Driving Laboratories for Chemistry and Materials Science" ↩

8. Google, "Google DeepMind and Isomorphic Labs introduce AlphaFold 3" (May 8, 2024) ↩

9. Arc Institute, "Evo 2" (February 2025) ↩

10. Google DeepMind, "Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry" (October 9, 2024) ↩