An academic team has built a multi-agent system that edits long-form video end-to-end — something existing automated tools largely cannot do.
Researchers at HKUDS released VideoAgent, an open-source framework that coordinates more than thirty specialized editing agents to handle everything from shot planning to cross-modal content retrieval. The system parses user intent to filter which tools are relevant, then uses what the authors call a textual-gradient graph to sequence those tools into a coherent editing pipeline. On the team's own VideoEdit benchmark and on public datasets, VideoAgent hit orchestration success rates between 87 and 95 percent while cutting API costs by 60 percent compared to baseline approaches. Human raters scored its output only 4 percent below videos edited by people.
Most automated video tools today are built for short clips and single tasks — trim here, subtitle there. VideoAgent's architecture matters because it attempts to solve the narrative coherence problem: keeping a long video consistent in tone, pacing, and visual logic across many edits. That is the gap between a tool that saves time on grunt work and one that could plausibly replace a junior editor.
The 4-percent gap from human-level quality sounds close, but benchmarks have a habit of flattering their own creators — VideoEdit was designed by the same team, which is worth keeping in mind before declaring the junior-editor era over.