This paper presents HTC‑Claw, a new software platform that adds planning and decision making to large‑scale computational materials studies. The platform is built on the OpenClaw framework. Its goal is to turn slow, error‑prone manual workflows into an automated loop that goes from a user’s intent to final reports.

The researchers created an agent‑based system that translates high‑level research goals into many parallel tasks. In this context “agents” are software components that interpret instructions, plan jobs, watch progress, and analyze results. Key features are an automatic task decomposer, a closed‑loop execution engine that triggers real‑time analysis and reporting, adaptive decision logic that can start new calculations based on intermediate results, and a modular scheduler that is separated from domain code.

At a high level the software has three layers. The user instruction layer accepts natural language queries (for example, “Evaluate the band gaps of all spinel structures” or “Search for corundum materials that retain metallic properties under a 2% strain”). The OpenClaw decision‑making layer runs intent parsing, task planning, job scheduling, result analysis, and dynamic workflow control. The high‑throughput computing platform layer performs the actual simulations, such as structural optimization, electronic structure and optical calculations, mechanical property evaluation, molecular dynamics, and machine‑learning‑assisted computations.

This work matters because many current tools focus only on submitting batches of jobs and then rely on humans for planning, error handling, and interpretation. By adding automatic planning, real‑time analysis, and iteration, HTC‑Claw aims to change the workflow from “submit and monitor” into a full “submit–monitor–analyze–report” loop. The authors position their design against existing frameworks such as AiiDA, FireWorks, QMflows and recent multi‑agent systems, and they report case studies showing end‑to‑end operation from user intent to reporting.