MedClaw: New AI Agent Harness for Long-Horizon Surgical Video Reasoning

A novel AI framework named MedClaw has been developed to tackle the complexities of analyzing lengthy surgical videos, which can span several minutes and necessitate temporal reasoning throughout various stages. The framework, outlined in an arXiv paper (ID 2608.14015), distinguishes reasoning from perception through a text-only orchestrator that determines what evidence to collect and generates an auditable sequence of tool commands. Sub-agents, utilizing frozen vision-language capabilities, carry out these commands on the video pixels, engaging in tasks like viewing, cropping, and frame inspection. Unlike current approaches, which either compress procedures or require extensive data, MedClaw enhances context evolution instead of merely adjusting weights, presenting a promising avenue for surgical video analysis. This paper was shared as a cross-type submission on arXiv.