Towards agentic science for advancing scientific discovery
Hongliang Xin, John R. Kitchin and Heather J. Kulik, “Towards agentic science for advancing scientific discovery”, Nature Machine Intelligence 7(9), 1373–1375 (2025). https://doi.org/10.1038/s42256-025-01110-x
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
This is a short Comment (perspective) rather than a research article, so it reports no new data or experiments; the summary below describes its arguments and the systems it points to.
Question
The Comment asks whether AI agents that reason, plan and act on digital and physical systems with partial or full autonomy constitute a new paradigm of scientific discovery, and what must be solved for them to be used responsibly (Xin et al. 2025). The authors place this “agentic science” after a sequence of earlier paradigms: empirical discovery by intuition and trial and error, theoretical frameworks such as Newtonian mechanics, multiscale and multi-physics simulation enabled by high-performance computing, and the current data-science paradigm of mining high-dimensional datasets.
Approach
The piece is an argument built from a brief history and a survey of examples. It traces AI agents from DENDRAL (1965), which inferred molecular structures with rule-based reasoning, and the ELIZA dialogue program (1966), through expert systems, probabilistic models and machine learning, to deep learning in the 2010s and today’s large language models. It then reviews current agentic frameworks, discusses where they help and where they do not, and works through technical limitations, evaluation, reproducibility and ethics.
Key findings
The authors identify independent agency as the defining feature of agentic AI, with multimodal LLMs at its core. Science adds a requirement beyond ordinary text: symbolic and structured content such as chemical formulas and equations needs specialized tokenization, embeddings, and external knowledge bases or ontologies. Agents gather new data through active learning and tools, from software to automated hardware, and interoperability is improving through the Model Context Protocol (a standard interface between LLMs and tools) and the Agent2Agent protocol for communication among agents.
Examples cited include Sakana AI’s AI Scientist, which runs the full cycle from ideation to manuscript writing and review; the FutureHouse platform, which composes specialized agents for literature synthesis, novelty detection and experiment planning in chemistry; LLaMat, a materials-focused LLM for generating valid crystal structures and extracting literature data; and Coscientist, which designs procedures and runs cloud laboratory equipment through APIs. The authors draw a distinction using A-Lab, an autonomous solid-state synthesis lab combining robotics, conventional ML and ab initio calculations: it uses an LLM to predict synthesis protocols but is not agentic, because the LLM is not inside a closed reasoning or control loop. LLaMP, which uses hierarchical reasoning-and-acting agents with retrieval-augmented generation over materials databases and simulations, is offered as a genuinely agentic case. Impact is expected to be largest in fields with well-structured problems, abundant data and existing automation, with chemistry and materials science the prime examples, and smaller in fields with scarce unified data or ill-defined questions, or where human judgment and ethics must govern, such as clinical diagnosis.
The main technical concern is hallucination, including “slopsquatting” (originally, malicious registration of software packages that LLMs invent), and its compounding over long multi-step tasks, which the METR benchmark shows advanced agents still struggle with. The authors recommend grounding agents in knowledge graphs of validated facts (reaction networks, gene ontologies, crystallographic databases). For evaluation they note proposed metrics such as pass@k (success in at least one of k attempts), passk (success in all k), step counts and shortest-path distance in reasoning graphs, along with the AFMBench benchmark for autonomous microscopy, but stress these are not standardized or general across domains. They liken prompt fragility to sensitivity to initial conditions in a dynamical system and call for verification agents or human review at decision points, with automatic halting when a plan proposes something implausible, such as a nonexistent compound or an unsafe experiment. Further needs include standard protocols, data formats and APIs, scalable compute (weighed against its energy cost), and transfer, self-supervised and reinforcement learning.
Significance
The Comment argues that agentic AI could strengthen rigor and reproducibility by finding contradictions and gaps in the literature, provided that publications report the model and version, representative prompts and agent dialogues, and that journals adopt reporting guidelines akin to those for experimental protocols. On ethics it warns that agents may favor experiments reinforcing prevailing trends, and recommends diversified prompts, training data that include negative and null results, inspectable reasoning, and humans kept in the loop for strategic oversight while agents take on routine, high-throughput work. It continues the paradigm framing of the earlier Comment on modeling in chemical engineering.