Published in Nature on May 19, 2026 (doi:10.1038/s41586-026-10652-y) by FutureHouse with collaborators at the University of Oxford and Fordham University. Ali E. Ghareeb and Benjamin Chang contributed equally; Andrew D. White, Michaela M. Hinks, and Samuel G. Rodriques jointly supervised.
The claim
The paper's stated gap: "Scientific discovery is driven by the iterative process of observation, hypothesis generation, experimentation and data analysis. Despite recent advancements in applying artificial intelligence (AI) to biology, no system has yet automated all these stages."
Robin is described as "a multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology," which "can generate hypotheses, propose experiments, interpret experimental results and generate updated hypotheses, achieving a semi-autonomous approach to scientific discovery."
The provenance claim is unusually specific and is what makes the paper citable on autonomy rather than only on capability: "All hypotheses, experimental directions, data analyses and data figures in the main text of this report were produced by Robin."
The system is "semi-autonomous" by design — humans conduct the physical experiments. The automation covers "the key intellectual steps of the scientific method… while coordinating with scientists throughout the experimental loop."
Architecture
Three specialized language agents: Crow and Falcon for concise and deep literature search respectively, and Finch for experimental data analysis. Given a target disease, Robin "automatically identified relevant in vitro assays that model key disease mechanisms and proposed specific drug candidates to evaluate in these experimental models"; researchers ran the experiments and returned the data; Robin analysed it and "interpreted the results of this analysis to generate a new round of therapeutic candidates."
The result
Applied to dry age-related macular degeneration — "the major cause of blindness in the developed world" — Robin "proposed enhancing retinal pigment epithelium phagocytosis as a therapeutic strategy, and identified and confirmed in vitro efficacy for ripasudil and KL001."
Ripasudil "is a clinically used Rho kinase inhibitor that, to our knowledge, has never previously been proposed for the treatment of dry age-related macular degeneration"; KL001 "represents, to our knowledge, a novel approach to phagocytosis enhancement in RPE cells." Robin then "proposed and analysed a follow-up RNA sequencing experiment, which revealed upregulation of ABCA1, which encodes a lipid efflux pump and represents a possible novel target."
The authors are careful about what has been established: the hypothesis "would of course require validation in a suitable disease model and ultimately in a randomized, placebo-controlled trial to confirm clinical validity." What the result shows is in vitro efficacy and a mechanistic lead, not a treatment.
What kind of discovery this is
The paper's account of Robin's advantage is specific, and is the part most transferable beyond biology. Robin works by "'combinatorial synthesis' (identifying non-obvious connections between disparate fields)," which "effectively targets 'low-hanging fruit' that human experts may overlook due to the compartmentalization of scientific knowledge."
The motivating evidence is the historical lag in drug repurposing, where "although insights often existed in scientific literature, only after a substantial lag did that knowledge crystallize into a new treatment." The paper's examples: dabrafenib's otoprotective effects, discovered a decade after its molecular action was characterized and by unbiased screening rather than inference; ketamine (22-year lag); leucovorin (5-year); KarXT (13-year). The argument is that a system able to hold literature across fields can close lags produced by human specialization — "LLMs can store and recall information on a wide variety of scientific topics and thus transcend the limitations of individual human knowledge."
The context offered for urgency: "US Food and Drug Administration approvals stagnating at approximately 50 novel drugs annually over the past decade."
Safeguards
The paper describes four guardrails, unusual for a capability paper to state at this length. The system "prioritizes candidates with established safety profiles and searches for known toxicities or off-target interactions." As a lab-in-the-loop system, its outputs "are treated as therapeutic hypotheses that must undergo standard pre-clinical validation, ensuring that any unintended toxicity is identified through traditional in vitro and in vivo safety filters before clinical translation." On dual use, Robin "utilizes 'off-the-shelf' LLMs that have undergone extensive safety alignment (via red-teaming and reinforcement learning from human feedback) to prevent the generation of malicious biological protocols" — placing reliance on the underlying models' alignment rather than on system-level controls. And "all queries to our platform are run through an LLM classifier that checks against unsafe topics."
Stated limitations
Robin "generates experimental outlines, [but] it does not yet produce precise, executable protocols," so human translation remains necessary. Finch "is also reliant on prompt engineering by domain experts to produce reliable analytical results," which qualifies the autonomy claim at the analysis stage. The results rely on "frontier LLMs available in early 2025," though the architecture is model-agnostic.
The authors also anticipate their own obsolescence in part: "the rapidly advancing baseline capabilities of general-purpose coding agents may achieve parity with specialized harnesses such as Finch on standard computational tasks," while arguing domain-specific architectures remain necessary for enforcing experimental constraints.
Relationships
- supports: AI for Science — one of the first systems to close the hypothesis–analysis loop with a validated novel candidate
- related: FutureHouse, Healthcare — AI Deployment, Agentic AI, AI Biosecurity, Automated Alignment Research