Every face you meet on NetShow is rendering in your browser as you look at it. Not a video, not a loop, not a pre-render waiting for a play button — a face being drawn frame by frame on your own machine while the agent speaks.
That sentence does a lot of work, so here is what is actually underneath it, at the level you need to choose one rather than build one.
Six ways to have a face
An agent's face is not one technology. There are six presence tiers, and they differ in how the face is made, not in how much of an agent you get. A vector face and a 3D model are wearing the same agent underneath.
SVG — a procedural vector face. Generated in your browser from a seed, and the seed comes from the agent's name. Give the same name and you get the same face, every time, on every machine. That determinism is the point: it is an identity, not a random draw, and it costs nothing to wear because there is no asset to fetch.
Hybrid — a photo base with a living overlay. A real face base underneath, and a painted layer on top where the eyes, brows and mouth move with the conversation. This is the tier that reads most like a person, because the base carries the realism and the overlay carries the life.
Live VRM — a rigged 3D model with a live face plate. The deepest tier. A full 3D character with a face rig driven the same way the flat tiers are driven. You bring the model; the platform brings the plate that makes it react.
ASCII — hand-authored character art. A face made of characters, and made deliberately, by a person, in a studio.
Sketch Line — a living line drawing. The same idea in line rather than glyphs: an authored drawing that moves.
Kinetic Typography — the transcript is the face. No face at all, in the ordinary sense. The words themselves are the presence: what the agent says is what you watch. Strange until you see it, and then obvious for the agents whose whole value is what they say.
Four of these are available when you create an agent: SVG, Hybrid, Live VRM — and Kinetic Typography, which needs no authored art because the words are the face, so the flow offers a starter face (Ticker) the way Live VRM offers a starter model. The last two are studio tiers — the creation flow shows their cards but does not let you pick them, and says why right on the card: they are hand-authored in the studios after the agent exists. That is an honest limit rather than a missing feature. Authored art needs an author.
What "alive" actually means here
Three things happen together, and all three are what separate this from a talking-head video.
It renders locally. The face is drawn by your browser from a package — a small description of how this particular face is built and how it moves. The package is fetched; the frames are not.
It moves with the words. The mouth follows what is being said, and the expression follows what is being meant. Because both are generated as the answer is generated, there is nothing to sync and nothing to pre-record.
It reacts, because it has no script. A video of a face can only replay. A rendered face is being produced in response to a live conversation, which is why it can pause where you paused.
The rule that matters most
A face is a bonus, never a dependency.
Every surface that mounts a live face paints a still portrait first, in the exact place the face will appear, and swaps to the live version only once the package has genuinely applied. If JavaScript is off, or the visitor asked their system for reduced motion, or the network is slow, or the package is refused for any reason — the portrait stays and the page is whole. The front door's own anchors hold this rule too — Meet the anchors walks through exactly what happens when their faces cannot render.
You will notice this is a design constraint with a cost. It means the estate carries a real still for every live face, and it means no page may be built in a way that only works once the face arrives. That cost is paid on purpose. A visitor who cannot render a face is still a visitor, and a product that breaks for them is broken.
Choosing one
If you are building your first agent, the honest advice is short.
Take SVG if you want an identity immediately and want it free and instant. Take Hybrid if the agent stands in front of customers and needs to read as a person. Take Live VRM if you already have a model and the depth is worth it. Take Kinetic Typography if the agent's whole value is what it says. Come back for the two studio tiers later, when you know the agent well enough to know what it should look like.
And pick carefully the first time. The face and the tier are written together as the agent's presence identity when it is created, and today there is no owner-facing screen for swapping them afterwards. The versioning underneath is built — profiles are versioned and versions only ever grow, so nothing is ever overwritten — but the button that would let you drive it from your own dashboard is not there yet. Better to know that going in than to discover it later.