Researchers from Tinycloud presented a case study on how their production multimodal video agent harness represents skills for selection. The harness uses two recurring representations: tool-skills that wrap external APIs or system tools, and workflow-skills that orchestrate these calls to produce deliverables. A six-task selection ablation showed that full autoload of skills selects the correct skill every time, while partial exposure can create lexical competition that suppresses correct selection. This finding is connected to recent retrieval-based skill-routing work at large scale. --- Why it matters: This study matters because it highlights the importance of skill representation in multimodal agents and how it affects task completion. The findings have implications for the development of more efficient and accurate skill-routing systems, particularly in scenarios where partial exposure can lead to lexical competition. Source: https://arxiv.org/abs/2608.20389