A multimodal agent perceives and acts on multiple input types—such as an image, a spoken instruction, a document, and a video frame—together toward a single goal. The key challenge is the engineering distance between three models that work individually and one agent that reasons across all three at once. Perception occurs through completely different paths: images become visual embeddings, video adds a time dimension, audio becomes a transcript or acoustic features where something gets lost in conversion, and structured text like headers, tables, and code blocks must be parsed before reasoning. These pipelines were optimized in isolation, often by different vendors, and multimodal architecture requires stitching their separate outputs into one representation for the reasoning layer. Reasoning across modalities is fundamentally a trust problem: when two sources disagree, something must decide which one wins. The right answer changes with the task—a compliance agent should weight a document over a voice note, while a field service agent should weight the photo over the ticket description.