What's New in Visuals
Visuals now tells you what is wrong before it spends your credits.
Runs used to fail at the model. You pressed Run, waited, and got back a code. Now the problems are caught up front and named in plain English. Alongside that, the Visuals Agent can finally see your runs, and every setting in the app lives behind one door.
The short version:
Bad inputs are caught before the run, not after you have paid for it
The Visuals Agent can see your runs, even ones it did not start
The Agent remembers how you work, and you can teach it your own skills
One Settings home for profile, billing, usage, personalization, memory and skills
ElevenLabs voiceovers work again
Lyric videos no longer jump on the last clip
Video generation retries elsewhere instead of failing your card
Runs tell you what is wrong before you spend credits
Before, a bad input was discovered by the model. You pressed Run, waited, and got back a code.
Now the Run button turns off and says what is blocking it, and every card with a problem carries an amber badge explaining what to fix. Hover the badge for the full list.
Audio that is too long is caught before the run. A 199 second clip fed to a model that accepts 15 used to come back as an unreadable error code after you had already waited. Now it tells you the clip is 199 seconds, the limit is 15, and to trim it.
Multi shot durations have to add up, and now they do. Kling V3 plays your shots back to back and rejects the whole request if the shot lengths do not equal the total, and it only says so after the run has started. The settings panel now shows the shot total and offers a one click fix that sets the card duration to match, and a mismatch is refused before any credits are spent.
Changing a model reconnects your cards. Picking a model that does not accept one of the inputs you have connected now remaps or disconnects those connections and tells you what it did, instead of leaving dead connections behind.
Veo renders at the resolution you paid for. Attaching reference images was quietly forcing renders back down to 720p while you were billed for the full 8 seconds.
The Visuals Agent can see your work
It sees your runs, including ones it did not start. Ask why a run failed, or what a card produced, and it can answer instead of guessing.
It looks at what it made. The Agent inspects generated images and frames before it suggests a change, rather than reasoning about them blind.
It re-runs only what needs it. Point it at what broke and it fixes and re-runs those cards, instead of running the whole flow again.
Attach any card to the chat. Drop a card into the conversation to talk about that one thing.
It answers with things you can act on. Replies, run progress and suggested next steps come back as cards, and you can edit a draft inline without leaving the chat.
Clicking on the canvas follows one rule now. A single click anywhere on a card selects it, including on the thumbnail and on the prompt box. Shift and Command both work for multi select. Double click on a card's image, video or text opens the full view. Double click anywhere else on the card opens the edit panel. A plain click in the prompt still puts you in the editor.
The Agent learns how you work
This is the part to actually read. Three things now live in Settings, and together they are the difference between explaining yourself every time and being understood on the first try.
Custom instructions
Under Personalization, one place to write down how you want the Agent to work. Your style, your rules, the things you find yourself saying in every conversation. It travels with you into every chat, so you stop repeating it.
Memory
The Agent remembers the things worth remembering: your brand, a character you keep coming back to, a decision you made, a constraint you work under, a preference about how you like things done. Some of it it picks up as you work, some of it comes from what you told us when you signed up, and you can write your own.
Every memory is visible under Memory in Settings. You can read them, edit them, and switch any single one off without deleting it. If you would rather it did not remember anything for a while, pause it.
Memory is scoped, which matters more than it sounds. A memory can apply to everything you do, to one project, or to one flow. The most specific one wins. So a rule you set for a single project quietly takes over from your general one while you are working in it, and stops applying the moment you leave.
Skills
A skill is a set of instructions the Agent reaches for when the work calls for it, so you are not writing the same brief twice.
It already comes with a library of them, grouped by what they are for:
End to end, for whole pieces: a music video, an ad, a visualizer, a release pack.
Directors, the roles it works through first: the brief, the scenes, the motion.
Craft, for one thing at a time: a prompt, a visual document, a clip, a spoken line, a style, a preset, a model choice.
You can switch any of those off if it is not how you work.
You can also write your own, under Skills. A skill has a name, a short line describing when to use it, the instructions themselves, keywords, and reference assets you want the Agent to look at. That description line is what it uses to decide when to reach for the skill, so write it for someone who has to choose. Your own skills sit alongside the built-in ones, and each one has its own on and off switch.
One place for your settings
Settings used to be scattered across the app. There is now a single Settings home covering profile, account, billing, usage, personalization, memory, skills, and API and MCP access. It opens as a page or as a dialog inside the app, and the old profile settings page takes you straight there.
Your company name is editable. It was captured when you signed up and then locked. Now you can change it.
The left menu no longer disappears on a tablet. There was a band of screen widths where it vanished entirely.
Video generation is steadier
Seedance generations now route across more than one provider automatically. An infrastructure failure at one of them retries somewhere else instead of failing your card, and a reference photo that gets blocked is retried on the provider that accepts it rather than dead ending.
When something does go wrong, the message is written for you. No provider names, no raw error codes, and it tells you what to do about it.
Also fixed
Lyric videos no longer jump on the last clip. The zoom on the final clip snapped back to its starting scale about half a second before the end, which read as the video fading out and then coming back in. It affected most projects. Fixed.
ElevenLabs voiceovers work again. The last release said this was still broken. It is fixed, alongside the OpenAI and Gemini voices that were already working.
New flow in the New menu uses what you selected. Selecting assets in My Library and choosing New flow created a blank flow and ignored the selection entirely. Only the bulk bar button ever used it.
The bulk Flow button gives you something you can actually edit. It used to create a flow and a project at the same time and drop you into the project, so your edits never reached the flow. From source flow then showed you the original, un-edited version.
The Agent composer keeps the cursor with your text on mobile. On phones the cursor drifted further from what you had typed with every character.
The starting points on Create read more clearly, and there is a new one for cutting a set of clips for the feed.
Plus assorted polish on menus, tooltips, loading states, empty states and dark mode.