MatOS turns materials-research capabilities into reusable assets, scoring 11.15% above the previous best agent on end-to-end research with the same base model
Synopsis
A team at Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence released MatOS, an agent operating system for materials research that registers data, knowledge, tools, skills, roles and workflows as scientific assets with declared dependencies and provenance, and runs long tasks as checked execution graphs. On ResearchClawBench it scores 11.15% above the previously best-performing agent with the same base model and extends the cost-score frontier. Eight research cases span screening more than 30 million candidate structures to developing new methods.
Interpretation
Capability extension, long-task tracking and experience reuse for materials agents move into one operating-system layer: six asset types, namely data, knowledge, tools, skills, roles and workflows, each carry an identifier, declared dependencies and provenance, so a new database, code or method joins by registering an asset rather than rebuilding the agent. Materials and chemistry agents have mostly been specialized multi-agent pipelines, where adding a code or an experimental stage often means new tool functions and pipeline changes; recovery mechanisms such as workflow restarts, agent checkpoints and automatic restart of failed calculations act on single jobs or saved states. On the 12 ResearchClawBench tasks, the GPT-5.4 configuration of MatOS scores 19.46 overall, 11.15% above the highest-scoring baseline MIRA at 17.51; the cheaper DeepSeek-V4-Flash configuration scores 19.20, 9.63% above MIRA.
Long-running tasks run as explicit execution graphs whose nodes fall into decision, execution, check and control groups, and graph snapshots, node states and artifacts persist independently of the model context, so tasks can be paused, resumed and traced. When an agent revises its plan mid-study, the system must also decide which earlier results remain valid, and repeated runs of the same agent can still differ; existing recovery handles neither the validity question after a plan change. Graph updates proposed by the agent or a researcher must pass structural, policy and compilation checks before being admitted and compiled into a new snapshot, otherwise the current graph keeps running; every proposal, check and decision is recorded.
After a task, evidence-driven asset evolution locates the assets responsible from execution traces and user feedback, makes focused edits, and adopts them only after structural checks and semantic review, with foundation-model parameters held fixed. Earlier experience learning retains reflections, insights, workflows or tool knowledge and was mostly developed on general-domain tasks; in materials research a flawed simulation protocol, failure diagnosis or validation criterion can invalidate later results, so reuse needs checking. When a pre-update baseline exists, the structural check compares error sets from the original and candidate states and requires that the edit introduce no new error; the semantic reviewer asks whether the revision addresses the original evidence, preserves applicability conditions and avoids harming related capabilities.
Perspective
The asset library was frozen throughout evaluation, so no benchmark task contributed to asset revisions; on MatTools and SciAgentGym all three systems use DeepSeek-V4-Pro as the base model, and on ResearchClawBench all baselines use GPT-5.4. The eight cases cover inorganic materials, solid electrolytes, liquid electrolytes, alloys, and an OpenFOAM-based reproduction of the CFDBench lid-driven cavity.
In the alloy melting search, none of the 286 evaluated compositions meets all four acceptance criteria; the highest-scoring composition Ag3.02Bi1Pb85Pt4Sn6.98 (wt.%) has a residual solid fraction of 0.142, above the 0.1 threshold, and these surrogate predictions still need thermodynamic calculations or measurements. Liquid-electrolyte Onsager conductivities fall below Nernst-Einstein estimates and experiment, with the largest gaps in E2 and E3. The evaluation sets for the training-data selection method were also used during selector development.
