MatHarness: Towards Long-Horizon Open-ended AutoResearch for Computational Materials Science

Materials science research is a non-linear, long-horizon process. Existing scientific agents, however, largely rely on specific task execution or predefined workflows, rendering them brittle in open-ended settings. To bridge this gap, we present MatHarness, an agent harness designed for open-ended, long-horizon computational materials research spanning the complete lifecycle from initial ideation to manuscript writing. MatHarness couples three mechanisms: execution robustness through pre-flight verification, expert-distilled skills, and just-in-time auditing; epistemic adaptability through persistent research artifacts and an evolving scientific mental model; and collective intelligence through a heterogeneous multi-agent round table under targeted human oversight. We evaluate MatHarness on adsorbate-induced surface segregation in Cu-Zn alloys, an open question in methanol-synthesis catalysis. Over a multi-week trajectory consuming roughly 10^3 GPU-hours (50 production DFT relaxations prescreened by 171 machine-learning-potential evaluations) the system revised its plan six times, recovered autonomously from failed calculations, and had an independent auditor overturn one of its own over-attributed claims before publication. Within the sampled metallic dilute-limit models, the validated DFT analysis supports an adsorbate-induced tendency toward greater surface Cu exposure at both composition endpoints, with facet- and coverage-dependent exceptions. These results were compiled into a complete manuscript with human involvement concentrated in specification, approval, and strategic steering gates.