MatOS proposes an operating-system-style layer for AI agents in materials research
The arXiv paper describes a system that organizes data, tools and workflows as reusable assets so language-model agents can run long, resumable research tasks.
In plain words
Researchers increasingly use AI agents to search literature, run simulations and drive experiments, but such agents struggle to stay on task over days and to improve from experience. MatOS is a proposed framework that stores research resources as reusable assets, tracks tasks so they can be paused and resumed, and updates those assets from past runs and user feedback.
What was announced
The authors, writing on arXiv, say materials research depends on scattered data, knowledge and simulation tools, and that a single study can span multi-day workflows across literature analysis, computation and experiment. While language-model agents can already retrieve materials data, operate simulation codes and work with autonomous laboratories, the paper identifies three persistent difficulties: adding new capabilities to agents, keeping long-running tasks on track, and letting agents learn from past tasks and user feedback. MatOS addresses these with three components: a scientific asset graph that treats data, tools, skills, roles and workflows as reusable assets; a stateful adaptive workflow that runs tasks as explicit execution graphs with scientific checks so they can be resumed and traced; and evidence-driven asset evolution, which revises assets based on execution records and feedback.
In the paper's reported evaluations, MatOS improved on the previously best-performing agent by 11.15% on ResearchClawBench using the same base model, and outperformed Codex CLI and Claude Code with a common base model, exceeding Codex CLI by 2.13% on MatTools and 4.17% on SciAgentGym. The authors also describe eight research cases, including narrowing more than 30 million candidate structures to 97 solid-electrolyte candidates and identifying 89 halide variants with higher computed conductivity than their parent structures. These are the authors' own reported results, and the abstract does not describe independent verification. The code is available on GitHub.
What it does not show
- The source is the paper's abstract only; full method details, evaluation setups and caveats are not available in the supplied passages.
- All benchmark scores and research-case outcomes are the authors' self-reported results; no independent confirmation is described.
For security and IT teams
- Access
- The authors state that MatOS code is available at a public GitHub repository.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.