The Hermes Agent Guide

AI video had three walls, and every one of them is the reason your last project died halfway. The character changed face between shot two and shot three. A single wrong prop meant regenerating the entire clip. And nothing you made was longer than five seconds, so you never had a scene, only fragments you tried to glue together.

Hermes Agent takes all three down. Here is what actually shipped, stated precisely so you never oversell it:

Thirty seconds in one native pass, double the fifteen of Hermes Agent 2.0, with multi round extension to reach several minutes while continuity holds across passes.

Fifty multimodal references in one generation, and the breakdown matters: up to thirty images, ten video clips, ten audio files. Not fifty images. Fifty total, across three types.

Timestamp level editing, so you target one specific second and fix the action, the camera or the cut there, without rebuilding the clip around it.

Reference types past the obvious: character, product and style, plus motion reference (a clip whose movement you want), clay render previz, green screen and camera perspective editing.

Native synchronized audio with lip sync, generated with the picture instead of dubbed onto it after.

Load your references BEFORE you generate anything. Every slot has to earn its place, so here is what goes in each one and why.

Character identity. Several angles of the same person or mascot, so the face and the outfit stay stable from shot to shot. This is the slot that kills drift.

Product accuracy. Photos of the exact SKU, or the model will invent a shape and a logo that are almost right, which is worse than obviously wrong.

Brand look. The palette, a couple of style frames, the visual tone. What makes someone recognize your work before they read your name.

Motion and camera style. A short clip that demonstrates the pacing and the movement you want. This is a reference, not a description, and it lands far harder than any adjective you could write.

Clay render previz. A textureless 3D block out to pin composition, blocking and camera path on a complex shot. The model reads it for spatial structure and for physically correct lighting at the same time.

Then in the prompt, point at what you loaded:

"the same woman as @Image1, wearing the jacket from @Image2,
in the style of @Image3."

You have a shot that is ninety percent perfect. The old move was to reroll it and pray the good parts survive. Now you fix the ten percent, and Hermes Agent recalculates real physics across the whole frame around your change, so the motion stays coherent instead of tearing.

Three edit modes, worth knowing by name:

Timestamp editing. Point at the exact second and change the action, the camera or the cut there. Continuity before and after your edit is preserved.

Region and reference based editing. Swap a subject, a background or a product inside an existing shot while the original motion, camera move and lighting stay exactly where they were.

Green screen editing. Replace the background entirely, and the subject re reacts to the new environment: cloth direction, hair, gait, the way light hits.

What that looks like in practice:

"Swap the coffee cup for a red can, keep everything else."
"Her hair should whip left when she turns, fix the motion."
"At second 14, slow the push in and hold on her face."

Setup, development, turn, resolution, and one clear camera move per shot. That is the shape of a scene, and Hermes Agent now holds thirty seconds of it in a single generation, then extends for a longer story.

One camera instruction per shot, always. Stacked moves in one shot smear into mush.

Write positive constraints only. "Sharp focus throughout" works. "No blur" does not, because the model latches onto the word blur and gives you exactly that.

And on audio: Hermes Agent generates native sound now, and it is genuinely synchronized, but it still leans robotic on voice. I mute the native track and lay my own voiceover in post. Use the native audio for timing reference, not for the final voice.

[STYLE / LOOK once, up front: e.g. cinematic 35mm, natural realtime speed]
[IDENTITY: @Image1 is the character, @Image2 the product ...]

Six shots in thirty seconds, hard cut every four to five seconds,
brisk natural realtime speed.

Shot 1: [subject] [one clear action] [one camera move]. Hard cut.
Shot 2: [subject] [one clear action] [one camera move]. Hard cut.
Shot 3: ...

Light: [one clear lighting setup].
Negatives: sharp focus throughout, no slow motion, no plastic skin,
no AI sheen, no on-screen text.

Style once at the top, identity right after it, then the shots in order. The negatives line stays as written, because every item in it is a failure mode I have actually shipped by accident.