Page 131 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 23 of 31

with automated robotic systems, it achieves end-to-end experiment validation without human intervention.
This demonstrates the capacity of LLM-driven agents to autonomously design, plan, and execute complex
experiments .
[134]

Furthermore, the introduction of multimodal perception and knowledge fusion strategies can overcome the
limitations of single-data-dimension approaches, equipping agents with advanced cognitive capabilities that
surpass those of traditional single-data-stream methods. For instance, Copilot for Real-world Experimental
Scientists (CRESt) introduces a multimodal model to integrate composition, text, and microscopy image
data. By leveraging multimodal models, knowledge-assisted Bayesian optimization, and robots, it screened
900 catalyst chemical compositions within three months, ultimately discovering a catalyst in eight-element
compositional space with a 9.3-fold performance improvement [135] . Meanwhile, Rainbow used parallel
multi-robot operations and a real-time spectral feedback loop to autonomously optimize the luminescence
properties of metal halide perovskite nanocrystals, achieving closed-loop exploration of high-dimensional
synthesis parameter spaces .
[136]

Multi-agent collaborative architectures can further enhance system complexity and adaptability, enabling AI
agent systems to handle complex research tasks. For instance, Tippy deploys five specialized agents working
collaboratively to achieve full automation of the design-synthesize-test-analyze cycle in drug discovery,
significantly improving workflow efficiency and decision-making speed [137] . ChemAgents employs a
hierarchical multi-agent structure to manage overall laboratory resources. Through a task manager, it
coordinates literature readers, experimental designers, computational executors, and robotic operators,
demonstrating powerful capabilities from simple syntheses to multi-step complex experiments [138] [Figure
7B]. An LLM-based reaction development framework (LLM-RDF) utilizes six specialized agents working in
concert to accomplish end-to-end synthesis development tasks from literature mining, experimental design,
robotic execution, and product purification [139] . MatPilot (an LLM-enabled AI materials scientist) [140] and
MatAgent [141] integrate human scientific intuition with AI’s high-dimensional information processing
through human-machine collaboration based on natural language to jointly drive materials discovery. The
Andrew I. Cooper team has enabled agents to share existing laboratory equipment with human researchers
through modular mobile robots, automated synthesis platforms, and characterization devices without the
need to redesign the experimental equipment, which greatly accelerates the development of agents using a
wide variety of experimental equipment for the synthesis of multiple types of materials . It is worth noting
[142]
that these systems have moved from laboratories to applications in specialized environments. For instance, a
robotic AI chemist can automatically process Martian meteorites, independently synthesize and optimize
oxygen-evolution catalysts, and quickly screen out efficient formulas from 3 million combinations, validating
the feasibility of autonomous laboratories in extreme environments .
[143]

Although significant progress has been made, there are still some key challenges in the deployment and
expansion of agent-driven SDLs. Firstly, data quality and standardization pose significant obstacles. The
diversity of experimental conditions, instrument calibrations, and laboratory data formats hinders the
training of robust and generalizable AI agents. Secondly, the integration with existing equipment and
heterogeneous hardware platforms brings technical and compatibility challenges. Not all laboratory
equipment is designed with robots or software control in mind, which makes full automation complex.
Finally, ethical considerations and safety assurances in autonomous experiments, such as handling hazardous
chemicals or ensuring repeatable results, require the establishment of protocols and continuous monitoring
mechanisms.

The evolution path of agent-driven SDLs progresses from modular automation equipment to cognitive
agents empowered by LLMs, then further enhanced through multimodal perception and multi-agent