Abstract
<title>Abstract</title> <p>GUI-based CAE software sits at the center of aerospace analysis workflows, yet interacting with it remains a labor-intensive bottleneck. Embedding foundation-model agents inside such software demands two capabilities that existing work lacks. The first is reliable tool calling over complex long-horizon chains, synchronized with the real software state. The second is self-evolution: tool surfaces and user habits drift over time, and the harness must improve with use. This paper presents HABIT (Harness with Adaptive Behavior Inferred from Trajectories), a selfevolving, host-agnostic harness for GUI-based CAE software. It builds on seven design principles, a five-layer architecture with context engineering, and a four-mode taxonomy of human-AI collaborative GUI control. A memory system passively learns stable habits from operation trajectories and personalizes behavior across sessions. HABIT is evaluated on a 33-task post-processing suite with final-state verification, and demonstrated end to end on the CRM-WBT wing-body aerodynamic benchmark, where a habit-driven session replays mined user preferences from a single instruction.</p>