AI Agents Close the Discovery Loop in Two Nature Studies

For decades, the promise of artificial intelligence in the laboratory was mostly about pattern-spotting: sift a mountain of data, flag the anomaly, hand it back to a human. Two papers published in Nature on June 30, 2026 mark a different kind of moment. Both describe systems in which teams of autonomous AI agents do something researchers once considered the irreducibly human part of science — proposing original hypotheses, designing the experiments to test them, and revising their thinking when the results come back. In one case, the loop ran all the way to a validated drug candidate for a leading cause of blindness.

The first paper, from the startup FutureHouse, introduces Robin, described as a multi-agent system for automating scientific discovery. The second, from a Google DeepMind team led by Juraj Gottweis, reports on Co-Scientist, a Gemini-based system that acts as a collaborative research partner. A companion News & Views article by Weill Cornell Medicine's Olivier Elemento framed the pair together, noting that they push toward a laboratory discovery cycle with AI involved at every step. Read side by side, they suggest the field is crossing from AI that assists thinking to AI that helps run the scientific method itself.

Why It Matters

The headline result belongs to Robin. Applied to dry age-related macular degeneration — the most common cause of vision loss in the developed world, and a condition with few effective treatments — the system proposed a therapeutic strategy of enhancing the way retinal pigment epithelium cells clear cellular debris through a process called phagocytosis. It then identified and, working with human scientists running the bench experiments, validated a promising candidate: ripasudil, a drug already approved for glaucoma in some markets. Robin went further, suggesting an untested mechanism involving the cells' circadian rhythm and an experimental compound, KL001, never before proposed for the disease.

That matters because the slowest, most expensive part of modern medicine is not manufacturing a drug but finding the right target and the right molecule in the first place. Drug discovery routinely takes more than a decade and billions of dollars, and most candidates fail. A system that can read across a fragmented literature, connect dots no single specialist would spot, and hand a lab a ranked, testable idea does not replace the years of trials that follow — but it compresses the part where good ideas are easiest to miss. Google's Co-Scientist made a parallel case in drug repurposing for liver fibrosis, where it recommended targeting epigenetic modifiers; two of three suggested drugs, including the FDA-approved cancer therapy vorinostat, showed meaningful anti-fibrotic activity in human liver organoids.

Inside the Systems

Neither system is a single model answering a prompt. Both are built as coalitions of specialized agents that argue with one another. Co-Scientist runs what its authors describe as a tournament of ideas: generation agents propose hypotheses, reflection and ranking agents critique and score them, and evolution agents refine the survivors over many rounds, all grounded against scientific literature and structured databases. Every claim is meant to arrive with a citation a human can click and check.

Robin closes a tighter loop. It pairs literature-search agents with data-analysis agents so the system can propose an experiment, take in the raw results once human collaborators have run it at the bench, interpret those results, and then generate the next round of hypotheses — a semi-autonomous cycle rather than a one-shot suggestion. The distinction is the point: earlier AI research assistants tended to stop at the idea. These systems are designed to keep going, feeding what the experiment revealed back into what to try next. The humans still hold the pipettes and the final judgment, but the reasoning scaffolding runs continuously in the background.

The Reaction

The scientific response has been a mix of genuine excitement and pointed caution. In an accompanying Nature correspondence published the same day, researchers made the case that AI tools can speed up thinking, but evidence still comes from the lab bench — a reminder that a hypothesis, however elegantly generated, is only as good as the wet-lab work that confirms or kills it. Robin's own results underscore this: ripasudil looks promising precisely because human experiments backed it up, not because an agent asserted it.

There are deeper worries too. Some scientists question whether systems trained on the existing literature will tend toward safe, incremental ideas rather than the genuinely surprising leaps that reshape a field — a concern that AI could nudge research toward a productive but narrow monoculture. Others point to reproducibility: an agent that generates thousands of plausible hypotheses can also generate thousands of plausible-looking dead ends, and telling them apart still demands expert human judgment and time. Elemento's commentary welcomed the progress while stressing that rigorous, independent validation has to remain the gatekeeper.

What Comes Next

The near-term trajectory points toward tighter integration between these reasoning systems and physical laboratories. FutureHouse and others are already gesturing at automated or self-driving labs, where robotic instruments carry out the experiments an agent designs, shrinking the delay between hypothesis and result. Google, meanwhile, is folding this line of work into a broader Gemini for Science effort and gradually opening access to research partners, alongside tools for literature synthesis and agent-assisted peer review.

The open questions are less about capability than about trust and access. Who is accountable when an AI-proposed experiment misfires? How do journals credit discoveries where an agent did much of the reasoning? Will these tools stay concentrated in a handful of well-funded labs, or reach the smaller groups that make up most of global science? The next year is likely to bring more real validations like ripasudil, but also the first serious efforts to write the norms — around disclosure, replication, and authorship — that decide how much weight a machine-generated hypothesis is allowed to carry.

Closing Thoughts

What makes these two papers feel like a threshold is not that AI proved smarter than scientists. It is that, for the first time in peer-reviewed form, autonomous systems participated in the full arc of discovery — question, experiment, interpretation, revision — and one of them helped surface a real candidate for an untreatable disease. The romance of science has always centered on the lone insight, the flash of intuition at the bench. These systems suggest a quieter future, where much of the connective reasoning is shared with tireless software partners, and the human role shifts toward asking the right questions and deciding which answers to trust.

That future is not here yet, and the caveats from the bench are worth taking seriously. But the direction is now visible in the literature rather than the press release. When the record of how a disease was finally treated includes a line about the reasoning an AI agent contributed, the question stops being whether machines can do science and becomes how we want them to.

한글 요약

2026년 6월 30일 네이처에 발표된 두 편의 논문이 과학 연구에서 인공지능의 역할이 한 단계 넘어섰음을 보여줍니다. 스타트업 퓨처하우스의 '로빈(Robin)'과 구글 딥마인드의 '코사이언티스트(Co-Scientist)'는 여러 자율 AI 에이전트가 팀을 이뤄 가설을 세우고, 실험을 설계하며, 결과를 해석해 다시 다음 가설로 이어가는 시스템입니다. 특히 로빈은 선진국 실명의 주요 원인인 건성 황반변성을 대상으로 망막색소상피 세포의 노폐물 제거 기능을 강화하는 전략을 제안하고, 녹내장 치료제 리파수딜을 유망 후보로 실제 검증하는 성과를 냈습니다.

두 시스템 모두 하나의 모델이 답하는 방식이 아니라, 가설을 생성·비판·평가·개선하는 특화된 에이전트들이 문헌과 데이터베이스에 근거해 토론하듯 작동합니다. 구글 코사이언티스트는 간 섬유화 치료제 재창출에서 후성유전 표적을 제안했고, 그중 FDA 승인 항암제 보리노스타트를 포함한 두 약물이 인간 간 오가노이드에서 유의미한 항섬유화 효과를 보였습니다. 신약 개발에서 가장 느리고 비싼 단계가 올바른 표적과 분자를 찾는 일이라는 점에서, 파편화된 문헌을 가로질러 검증 가능한 아이디어를 제시하는 능력은 의미가 큽니다.

다만 같은 날 실린 네이처 서신은 "AI가 사고를 가속할 수는 있어도 증거는 여전히 실험대에서 나온다"고 지적했습니다. 기존 문헌으로 학습한 시스템이 안전하고 점진적인 아이디어에 치우칠 수 있다는 우려, 재현성 문제, 그리고 소수의 자금력 있는 연구실에 집중될 가능성 등도 과제로 남습니다. 앞으로는 로봇 실험 장비와 결합한 자율 실험실, 그리고 공개·재현·저자 표기에 관한 규범 마련이 관건이 될 전망입니다.