Explainer: How 'Agentic' AI Systems Are Automating Scientific Discovery and Lab Work

For decades, the ultimate running joke in artificial intelligence research was that scientists would eventually build a model capable of writing their own research papers for them. In 2026, that punchline has quietly become a reality. Across the globe, a new class of "agentic" AI systems is moving beyond simply answering questions or generating code snippets. These systems are now autonomously designing experiments, running complex simulations, and even operating robotic wet labs. This marks a profound shift in how scientific discovery is conducted, transitioning AI from a passive computational tool into an active, independent research collaborator.[2]

The core difference between a standard large language model and an AI "agent" lies in autonomy and feedback loops. Traditional models require a human to prompt them at every step. Agentic systems, by contrast, are given a high-level goal—such as "find a compound that binds to this liver disease protein"—and are equipped with a suite of digital tools to achieve it. They can search existing scientific literature, formulate a novel hypothesis, write the Python code to test that hypothesis, and analyze the resulting data. If an experiment fails, the agent reads the error logs, adjusts its approach, and tries again without human intervention.

The most striking demonstration of this end-to-end autonomy comes from Tokyo-based Sakana AI, which recently unveiled a system aptly named "The AI Scientist." Developed in collaboration with researchers from Oxford and the University of British Columbia, the framework automates the entire lifecycle of machine learning research. Given a broad direction and a starting codebase, the system brainstorms ideas, executes the necessary coding and experiments, visualizes the results, and writes a complete scientific manuscript.[1]

Remarkably, Sakana AI reports that the computational cost to produce a full research paper through this automated pipeline is approximately $15. The system even includes an automated peer-review module that critiques its own generated papers with near-human accuracy, allowing it to iteratively improve its findings before finalizing the manuscript. While its current focus is on computer science and machine learning, the underlying architecture proves that an AI can successfully navigate the open-ended, iterative nature of the scientific method.[1]

But the ambition of agentic AI extends far beyond writing code and publishing papers; it is actively crossing over into the physical world of biology and chemistry. At Stanford University, researchers are utilizing an agent called Biomni, which integrates hundreds of specialized databases and software packages into a unified environment. Stanford scientists can prompt the agent to design a wet-lab experiment or review biomedical data, and the model autonomously executes the complex workloads required to generate new biological hypotheses.

This integration of AI and physical laboratory work is rapidly commercializing. Insilico Medicine recently launched LabClaw, an intelligent laboratory operating system that bridges the gap between computational prediction and physical execution. Described as an "Agent-Guard" architecture, LabClaw utilizes five collaborative AI agents and 28 specialized skill modules to manage an end-to-end loop of drug discovery.

Instead of a human scientist manually pipetting samples or coordinating equipment schedules, LabClaw autonomously manages target selection, reagent preparation, and data analysis. The system incorporates a "human-in-the-loop" checkpoint at critical junctures to ensure safety and compliance, but the bulk of the tedious orchestration is handled by the machine. By taking over these standardized operations, the AI frees human researchers to focus entirely on high-level scientific judgment and creative problem-solving.

The impact on research velocity is staggering. In early deployments of similar agentic systems, research teams have reported slashing the time required for certain drug discovery phases by up to 200-fold compared to human scientists working alone. DeepMind's Co-Scientist, for example, was able to identify existing drugs that could be repurposed for a specific type of leukemia within a matter of hours, a process that traditionally takes months of manual literature review and cross-referencing.

To support this ecosystem, major infrastructure providers are building the foundational tools required for agentic biology. NVIDIA's BioNeMo toolkit now allows developers to transform general-purpose AI agents into specialized life sciences assistants in a matter of minutes. By providing access to a decade's worth of biomolecular libraries and open models, BioNeMo equips these agents with the ability to predict protein structures, run generative chemistry simulations, and execute genomic analyses with the speed of a supercomputer.

Beyond biology, agentic AI is tackling the fundamental laws of physics and national security challenges. At Los Alamos National Laboratory, scientists have developed the Universal Research and Scientific Agent, or URSA. Designed to operate in highly complex domains like radiation-hydrodynamics, URSA acts as a team of specialized agents that brainstorm hypotheses and run intricate simulations.

What makes URSA unique is its grounding in physical reality. Rather than just generating text, the system reasons using mathematical equations, physical models, and real-world experimental data. This allows the AI to navigate massive design spaces and find optimal solutions for fusion energy or materials science much faster than traditional linear workflows. By adjusting its strategies based on intermediate simulation results, URSA handles the inherent uncertainty of cutting-edge physics research.

Despite these breathtaking advances, the technology is not without significant limitations and risks. Current AI scientists lack physical intuition and "vision"; they cannot easily tell if a generated data plot is visually unreadable or if a physical lab setup is fundamentally flawed. Furthermore, these models can still hallucinate, occasionally making basic mathematical errors or misinterpreting the results of their own code.[2]

There are also profound safety concerns regarding emergent behaviors. As AI agents are granted more autonomy to write code, run experiments, and optimize their own objective functions, the risk of unintended consequences grows. If an agent is tasked with finding a novel chemical compound, it must be rigorously constrained to ensure it does not inadvertently design a toxic or hazardous substance in its pursuit of optimization.[1][2]

To mitigate these risks, the current consensus across the industry is to treat AI agents as "co-pilots" rather than fully autonomous replacements. The human scientist's role is evolving from a hands-on laboratory technician into a creative director. Humans set the ethical boundaries, define the overarching research goals, and provide the critical real-world intuition that machines still lack.[2]

As these agentic systems continue to scale, they promise to democratize access to top-tier scientific capabilities. A small biotech startup or an independent university lab equipped with an AI scientist could soon wield the research output of a massive pharmaceutical conglomerate. By automating the most labor-intensive aspects of the scientific method, humanity is poised to accelerate its understanding of everything from chronic disease to the fundamental nature of the universe.