Agentic video editing · field-test → shipped

The typographic social cut,
now reachable on-network

A creative-agency field test handed an agent a real social-editing brief — captions that hit every word, an overlay clip that drops in on a beat, a16z-style molecular graphics, a grade. The network couldn't reach it, and the job shipped at $0.00 network spend. Four capability fixes closed the gap. This is the brief, re-run end-to-end as four reproducible tool calls.

Source brief — a16z "agentic editing", GTM kickoff cut Layers — captions · PiP · typography · grade Network spend — was $0.00 · now ~$0.001/s per finishing layer Live on the network
Composited frame: condensed hero type, molecular network graphics, word-hit captions over a graded talking-head
One hyperframes-render pass. Condensed hero type (brand word in yellow), an a16z-style molecular network, and word-hit caption slabs — pixel-perfect HTML/CSS composited over the footage, then graded.
Same frame during the 1.5–3.5s window with a timed picture-in-picture in the upper right
Timed picture-in-picture. The overlay clip (upper-right, x:0.6 y:0.04) appears only in the 1.5–3.5s window — the "drops in when I raise my hands" beat. One overlay call, no scene-slicing.
The typography layer, moving. Kinetic word-hit captions + floating molecular graphics, rendered from HTML/CSS over the talking-head. The only listed primitive that could produce on-video type — and, until RC-1, undiscoverable.

Four layers, four calls

1

Word-level captions RC-3

transcribe(granularity:"word", burn:true, font_size, font_color, outline) — word granularity routes ASR to the live whisper-word cap for per-word timings (the "hit every word" effect), with fill / stroke / size / position forwarded to the burn. No more local faster-whisper.

2

On-video typography + graphics RC-1

hyperframes-render(background_video_url, html) — the condensed type, kinetic captions, and molecular-network SVG are ordinary HTML/CSS composited over the footage. RC-1 published its input schema, so an agent can call it first-try instead of a probing round-trip.

3

Timed, positioned overlay RC-2

overlay(image_url, x, y, scale, start_sec, end_sec) — a picture-in-picture at any x/y, visible only in a time window. Before RC-2 this forced slicing a 17s brief into 14 scenes; now it's one call.

4

Finishing colour grade RC-4

ffmpeg-colorgrade(preset:"cinematic") — a named look or explicit brightness/contrast/saturation/gamma. The finishing grade the tool caps were missing — now a real cap, stated rather than discovered.

The finding, not a saving. "Zero network spend" wasn't a cost win — it was one contract failure wearing several masks. Almost every primitive the brief needed already existed as compute; the agent just couldn't reach it because the capability cards didn't publish enough to plan or call. The fix was mostly contract, not compute. Full write-up: GTM editing-gaps analysis & proposal.

Make it yours

The same four calls cover a wide family of cuts — swap the design system, the overlays, the grade. The repeatable, parameterized recipe is the playbook:

Use caseWhat changes
Neutral / UI design systemSentence-case type, weight-not-colour captions, no decoration, footage is the only colour — pass a design_ref and it's honored.
Product / e-commerceProduct b-roll + a price/spec lower-third + a detail-crop PiP at the reveal + a vibrant grade.
Localized versionsRe-run the caption step per language (auto-detected word timings); keep the typography layer.
Brand-locked repeatabilityBind a brand kit (palette + finishing chain) so every cut matches.

▸ Open the playbook — fill in the brief and run it against your own footage.

Rendered on the Livepeer network · chatterbox-tts → talking-head → whisper-word/transcribe → hyperframes-render → ffmpeg-overlay → ffmpeg-colorgrade. Born from the a16z agentic-editing field test that RC-1…RC-4 closed.