waitless

A few thoughts on agentic video editing after 3M longform YouTube views, edited using agents: 1. Agentically edited content has a "look": it's heavy on motion graphics (since they are easy to make with code), lots of text, heavily animated. To stand out, we've been spending lots of time making our system do things that make it not look like that (custom shaders, imperfection added on purpose, ..) 1. In 1 year 90% of videos will be edited fully agentically, the remaining will use agentic bits with some human editing to polish 2. Making one video is easy, building a system and process that scales to multiple ones is much harder. We built a harness for claude that analyzes script, creates storyboards in Paper to iterate on and animates dozens of animations per video at scale, with human feedback on Flask. 3. One person with their agent is easy. As soon as you have a team, it gets 10x harder because you need to share code. GitHub is too complex and not built for this. We use Flask .do (frame io for agents) to store video, code and context into each version so everyone can just pull from there 4. MCPs for traditional tools like DaVinci, After Effects are slow and token intensive and in most cases are not worth it. Headless is the future, the hard part remains the tweaking and that's still unsolved