Integrating multimodal foundation models into enterprise ecosystems presents a fundamental software architecture challenge. Architects must balance competing quality attributes: the high latency and non-determinism of vision language action (VLA) models versus the strict determinism and real-time performance required by enterprise control loops. In this study, we propose an architectural pattern language for visual agents that separates fast, deterministic reflexes from slow, probabilistic supervision. It consists of four architectural design patterns: (1) Hybrid Affordance Integration, (2) Adaptive Visual Anchoring, (3) Visual Hierarchy Synthesis, and (4) Semantic Scene Graph.
@article{arxiv.2604.28001,
title = {A Pattern Language for Resilient Visual Agents},
author = {Habtom Kahsay Gidey and Alexander Lenz and Alois Knoll},
journal= {arXiv preprint arXiv:2604.28001},
year = {2026}
}
Comments
Accepted to the 23rd International Conference on Software Architecture (ICSA 2026), New and Emerging Ideas Track. 5 pages, 1 figure