Video components.

Prompt types into the real chat window, sends, reply streams in

ChatGPT

Claude

Gemini

Only the input box floating on the stage: prompt types, attachment chips drop in, send pulses

ChatGPT

Claude

Grok

Model/effort dropdown opens, cursor picks, it locks in

Agent CLI: task typed, Read/Edit/Bash tool calls tick by, diff summary, done line

Desktop agent app: thread list, active task streaming, status chips go running to done

Lock-screen / Mac notification from a personal agent, then push-in

You text your agent in a messaging app; typing dots; it replies with a done card

Agent cursor clicks through a website; step chips stack

Files changed, green/red lines, +312 -48 count up, merged badge

Chat on the left; the built thing slides out on the right and comes alive

Answer builds line by line while numbered source chips pop in

Agent pauses at a permission prompt; approve, it continues

The real usage panel: 5-hour and weekly bars fill to a percent

Call screen with your agent: live waveform and captions of what it is doing

Prompt card, arrow, the output under the product wordmark

The full prompt, big and readable; key phrases light up as read

The output builds up in stages while an elapsed timer runs, ending on a quiet '1 prompt · 16 min' stamp

Agent screenshots its work, rings the bug, fixes it, ticks it

Honest failure call-out on the output: strike, ring or red error line

A clean grid or reel of real things people built with a model, each with its creator handle and the model's mark

Local preview, URL types in, live site opens, link copies

The coordinator agent's live task list: parallel sub-agents appear as rows, each streams its status and ticks done, then a merged summary lands (a real product surface, not a node diagram)

A task passing between real surfaces: you ask your personal agent, it opens the job in the coding agent, and the result comes back to you (the actual chat, task and notification surfaces, not boxes and arrows)

The agent's real Connectors settings: Gmail, Slack, Notion, GitHub and Calendar switch on one by one and show connected

The laptop lid closes, the cloud task card keeps progressing, and a done notification arrives

Week strip with recurring task pills firing on time

Inbox sorts itself; Reply/Skip/Urgent chips land; drafted reply slides out

Agent cursor with a badge opens and drives a real desktop app

A SKILL.md card drops into the agent skills folder; a new ability lights up

The real memory surface: facts from a chat get saved into the memory list, then recalled in a new chat

3-5 models bars grow with logos at the base; the winner takes the accent

Old score, arrow, new score, delta counts up

A lab's model lineup as one clean spec picker: tiers side by side with speed, price and intelligence; one is marked recommended

Model versions on a dated line, each with its key number

Spec card: name, context, price in/out, date, availability chips

A container fills with documents up to the model limit

A sentence splits into token chips while a counter climbs

The new model slides into #1 and the rest shift down

One prompt into two real app windows; timers run; finish order lands

Two results side by side under their wordmarks; the winner ticks, the loser dims

Columns of price, context, speed, best-for with ticks

A long manual checklist of steps and tools collapses into a single prompt bar

A clean recommendation table: task rows, with the recommended model chip landing in each row

Lab mark, model name, version number ticks into place

Official post or system card scrolls; one sentence highlights

The real X launch post recreated with its media playing

Who gets it: tiers, regions, platforms light up

Subscription tiers; the one you need lifts

$ in / $ out per 1M tokens with the old price struck through

A receipt prints for one task: tokens, minutes, tool calls, total cost

Speaker/guest name and role over A-roll; a tasteful line, pill or accent bar that slides in and settles. AI context: the guest is an AI researcher; the name is a placeholder, e.g. "Your Name · Research Scientist, Agents"

Small corner credit for third-party AI footage, e.g. "SOURCE: @OpenAI" with the real favicon/mark

One to three floating keyword pills near the speaker's gesture, e.g. "Agents", "MCP", "Context window"

Corner badge counting through "4 reasons Claude Code beat Codex this week"

A sarcastic quoted claim beside the speaker: "AGI", "up to 10x faster", "PhD-level"

Date / relative-time context for a clip or launch, e.g. "Sep 29, 2026 · DevDay", "3 days ago"

Crediting whoever built the demo on screen: an X handle, a GitHub repo, a YouTube channel

A third-party AI clip (official launch footage, a screen recording of Claude Code) in a clean frame with a source line on the stage

Rounded presenter inset that keeps the face on screen while a full-frame AI graphic (a ChatGPT screen) plays

An AI product clip or screenshot centred over a blurred, darkened copy of your own footage

Two clips side by side with a divider and labels: ChatGPT vs Claude running the same demo; then vs now

A vertical phone capture of an AI app (ChatGPT voice, Grok, Gemini Live) centred in a bezel with clean sides

A screen recording of an AI tool with a smooth punch-in to one area and a ring on the target (the model picker, a tool call)

Your past AI video as a YouTube watch-page card: "remember when I tested GPT-6…"

A real-style launch or reaction post with the key phrase highlighted and optional media

A system card, research paper or PDF page: the rest dims, one line highlights, slow push-in

A screenshot of an AI product on the stage with a soft shadow and a source chip

An email (From / Subject / body) with the key line highlighted: a launch announcement, an agent's drafted reply

The usage bar hits the cap and the product's real limit banner slides in with a reset countdown