Contents
The prompt is Agent's job, not yours
Prompting was its own craft for a long time: to get a good frame, "make a clip" isn't enough — you had to know how to describe the scene for a specific model, where lighting matters, where camera movement does, where timecodes for shot changes go. Agent takes that craft on itself. You say the idea in human words — it composes a tuned, professional prompt: for the right model, with the right structure and the details that usually separate "meh" from "wow."
And you don't even have to know exactly what you want. Open the Agent tab in a project and just start talking: "I want a short clip about a coffee shop, but I don't know where to start." Agent offers a few concrete scene options, argues the details, helps build a shot list — and then turns that same conversation into a prompt and a generation. Discussing the idea and making it happen live in one window.
While Agent is working on a reply, you can see exactly what it's doing right now — not a summary after the fact, but step by step: "Thinking through the request," "Searching materials," "Picking a model," "Composing a prompt" — all the way to the finished result.
A proposal card: model, settings, and price — before you run it
When Agent is ready to propose a generation, a PROPOSAL card appears in the chat: the model and key settings, a prompt excerpt (expandable to the full text), and the exact cost in tokens — not an estimate range. If you have a Membership discount, the card shows the strikethrough base price next to the discounted total.
From there it's Generate or Reject. Until you hit Generate, the generation doesn't run and nothing is charged for it. The Agent reply that picked the model and drafted the prompt is paid separately, like any of its replies.
From launch to result — without leaving the chat
After you approve, the generation runs the same way it would from the editor: the same queue, the same billing, the same engine. Meanwhile the chat shows a live status — "Generating…" — and once the asset is ready, an inline preview appears with a link straight to Materials.
If a provider rejects the generation or something breaks, Agent doesn't just say it failed. It offers "Fix & retry": it works out what went wrong, rewrites the prompt (removing whatever might have triggered a block, or simplifying the request), and proposes a fresh card — with a new approval, not an automatic retry on your dime.
Your choice of approval mode and Agent model
By default Agent runs in Confirm mode: every launch is a card with a price and an explicit Generate. If pausing on every detail isn't what you want, switch to Auto (runs without pausing, up to a daily budget) or Hybrid (auto-runs cheap generations, still shows a card for expensive ones).
You can also pick which model drives the conversation itself — Agent isn't just Claude. The picker has Claude, ChatGPT, Grok, and Google models: powerful ones for complex briefs and fine edits, fast ones for quick, simple turns. The price of a reply depends on the model and the actual amount of work. Some models reply for free, and the picker shows their price as 0 tk.
Agent remembers your project — and you can run several chats
In "Project context" you can save notes by category — Rules, Brand, Personality, Style, Habits, General. For example: "the hero is a ginger cat named Mars" or "no neon." Agent factors these notes into every turn, so you don't have to re-explain context each time.
You can keep more than one chat per project — "New chat" starts another thread with its own title (auto-generated from the first message, but renameable). The chat list is searchable right there, and a link to a specific chat is preserved in the URL.
What's coming next for Agent
Part of this plan has already shipped: photos can be attached right in the chat (with brush-selected areas for targeted edits), Agent generates music and voiceover, and projects now have employees — eight specialists across content, strategy, and storytelling. See the "Agent 2.0: your project now has employees" announcement for details.
Still on the roadmap:
Skill management
Turn Agent's knowledge packs on and off by hand, for a specific task.
Employees working together
Specialists will learn to hand tasks to each other — from a script all the way to a finished clip with a soundtrack.
Frequently asked questions about Agent
Do I need to know how to write prompts?
No. That's the whole point of Agent: you describe the idea in plain words, and it composes the professional prompt for the right model itself — with all the structure, lighting, and details you'd otherwise tune by hand. If you do want the prompt as text, ask for "just the prompt" — but that's optional.
What if I don't have a clear idea yet?
Then talk it through in the chat. Tell Agent roughly what you want — it offers a few concrete scene options and helps build a shot list and mood. When the idea is ready, ask it to make it: Agent turns the conversation into a prompt and a generation, without switching to another window.
Is this the same as "Help with the prompt" in the generation bar?
No. "Help with the prompt" improves a prompt for the model and mode you've already picked. Agent is a separate chat: it picks the model itself, explains why, writes the prompt, and — with your approval — runs the generation, or just helps talk through an idea with no generation at all.
Do I get charged tokens just for asking for advice?
It depends on the model. A reply is charged for the actual work, and some models reply for free. To see the amount, hover over the reply on a computer or tap it on a phone. With a paid subscription, replies are paid first from daily Bonus tokens: 1 to 10 tokens a day, depending on your plan. In Confirm mode, generation itself is charged only after you press Generate on a card with the price.
Can I send Agent a photo?
Yes. Attach an image right in the chat — as a reference or for editing; you can brush-select the exact area to change and ask Agent to redo only that. Materials already uploaded to the project are available through the paperclip as well.
Does Agent only run on Claude models?
No. The picker has Claude, ChatGPT, Grok, and Google models. You choose the model right in the message field.
What happens if a generation fails?
Agent shows the error right in the chat, explains the cause, and offers "Fix & retry" — it rewrites the prompt and proposes a new run instead of a blind retry.
Can Agent generate music, voiceover, or avatars?
Music and voiceover — yes: Agent proposes them through the same flow as images and video, with a price card and your "Generate" before launch. Voiceover reads your text verbatim. Avatars still live in their own Studio.
Where are my chats with Agent stored?
Each project has its own chat list, with search and renaming. You can run several chats per project in parallel and switch between them.
