The complexity of modern materials research poses significant challenges to traditional “trial-and-error” methods and fragmented computational simulations, often resulting in low efficiency and high resource consumption. The emergence of artificial intelligence, especially large language models (LLMs), has brought transformative opportunities to computational materials science. Among these, multi-agent systems (MASs), which simulate the collaborative dynamics of human research teams, demonstrate exceptional promise for managing complex, multi-step scientific tasks.

This paper systematically reviews the applications of MAS in physical materials computation. We analyze how specialized MAS frameworks can achieve end-to-end automation of complex workflows, from initial structure modeling and parameter configuration to final simulation execution and result analysis. A detailed focus is placed on two pioneering systems: VASPilot and PhysAgent. The VASPilot is built on the CrewAI framework, and specifically designed to automate the entire workflow of the widely-used VASP (Vienna ab initio Simulation Package) software. It deploys a team of specialized agents (such as structural, calculational, validation agents) to streamline tasks such as band structure calculations, lattice optimization, and convergence testing, greatly reducing manual intervention and improving reproducibility. PhysAgent adopts a more ambitious “Mentor-Student-Leader” cognitive architecture, aiming not only at task execution but also at autonomous scientific discovery. It has demonstrated the ability to autonomously rediscover fundamental physical laws from simulated observational data, such as Kepler’s laws of planetary motion and Newton’s second law. Furthermore, PhysAgent successfully automates complex workflows such as the first-principles electronic structure calculations of GaAs and the simulations of complex phenomena such as raindrop flow on high-speed train windows.

Our analysis demonstrates that the MAS can substantially enhance computational efficiency, standardize procedures, and handle complex, non-routine problems beyond the scope of single-agent systems. The Key physical results and capabilities highlighted include the automated convergence testing of plane-wave cutoff energy (ENCUT) and k-point grids, comparative studies of van der Waals correction methods on material properties, high-throughput screening of transition metal dichalcogenides, and the autonomous derivation of physical laws from data.

However, through critical evaluation and practical case studies (such as using the SciLink framework for complex surface adsorption and defect calculations), we identify persisting limitations. These limitations include challenges in computational resource allocation, ensuring physical consistency in agent decisions, handling “parameter gaps” in multi-scale simulations, managing long-time-scale tasks, and a general lack of true scientific creativity for unsolved problems. The “black-box” nature of agent reasoning also hinders deep scientific trust.

Looking forward, we envision MAS evolving from efficient assistants to genuine scientific partners. This evolution will require 1) deeper integration of physical principles and constraints into agent reasoning, 2) the development of closed-loop research ecosystems including “hypothesis-design-calculation-experiment” cycles, potentially leading to self-driving laboratories, 3) expansion into frontier areas like catalytic reaction exploration, inverse materials design, and automated cross-scale modeling, and 4) fostering new researcher skills, with focusing on strategic problem formulation and critical analysis of AI-generated outputs. By addressing current challenges and leveraging these future directions, MAS hold the potential to fundamentally reshape the research paradigm in computational materials science and physics, accelerating the discovery and understanding of novel materials.