A multi-agent system architecture that parses clinical notes, synthesizes personalized exercise videos, estimates pose in real time, and delivers corrective feedback was detailed in an April 2026 arXiv preprint (Dharmaratnakar, 2026).

The prototype employs four micro-agents: Clinical Extraction via LLMs to derive kinematic constraints from unstructured notes, Video Synthesis built on foundational video generation models, Vision Processing with MediaPipe for pose estimation, and Diagnostic Feedback to issue instructions, according to the described pipeline (Dharmaratnakar, 2026; Wu et al., 2023). This directly targets reported physiotherapy compliance rates below 50% caused by absent personalized supervision, a gap prior static-video and generic-avatar systems failed to close (Jordan et al., 2021).

Original abstract omits error-propagation risks across chained agents and regulatory pathways for clinical deployment, elements examined in related multi-agent healthcare simulations (Li et al., 2024). The work synthesizes MediaPipe pose tracking advances (Grishchenko et al., 2020) with AutoGen-style agent conversation patterns (Wu et al., 2023), revealing a concrete transition of agentic AI into high-stakes rehabilitation where generative media supplies patient-specific training data unavailable in earlier computer-vision rehab tools.