Agentic screen recording in the browser

agent-video is an innovative system designed to create engaging, narrated screen recordings by mimicking a film production workflow. It operates through three distinct phases: pre-production, production, and post-production. During pre-production, the system uses an AI agent to visit webpages, capture accessibility snapshots, and generate structured narration with specific scroll targets, which is then synthesized into audio with precise timing data. The production phase involves recording the screen with content-aware scrolling synchronized to the narration. Finally, in post-production, relevant video segments are extracted, synchronized with the audio, and assembled into a polished video. This process allows for accurate, content-driven narration that matches the visual content precisely, eliminating timing drift and dead time, and providing a more engaging viewing experience. The system supports customizable personas for narration, making it versatile for various applications such as product demos, competitor analysis, bug reports, and more, by leveraging existing technologies like language models, text-to-speech, and video hosting in a novel, integrated manner.