How AI agents actually edit your timeline

Most “AI video” tools work like vending machines: prompt in, finished MP4 out. If the result is 90% right, there's no way to fix the last 10%, you re-roll and hope. That's a demo, not a workflow.

We took the opposite approach. In Agent28, the agent's output isn't a video file, it's a sequence of timeline operations. Every cut, caption, zoom, and overlay lands on the timeline as a normal, editable object.

Step one: understand the footage

When you upload footage, we build a machine-readable model of it. Vision models describe what's on screen, shot by shot. Word-level transcription aligns everything said to the exact frame it's said in. Audio analysis marks energy, silences, and beats.

Step two: plan like an editor

Your prompt is interpreted against that model. “Cut the rambling in the intro” becomes: find low-information segments in the first section of the transcript, cut them, and heal the pacing around the cuts. “Punch in on the reveal” becomes: locate the reveal visually, add a scale keyframe pair at the right beat.

The agent plans in the same vocabulary an editor thinks in: shots, beats, pacing, emphasis.

Step three: execute transparently

The plan compiles down to timeline operations, and each one is applied where you can see it. Nothing is flattened. If a caption is two frames late, you drag it. If a cut feels aggressive, you pull the edge. Or you tell the agent and it does another pass.

This is the property we care most about: the agent never takes your project somewhere you can't follow. It's an editor working in your project, not a generator replacing it.