A full live demo of the in-canvas design agent: connectors, custom skills, and kicking off work in parallel
Figma deep dive: Agents | Config 2026 · Angel
27 min total·Actually worth watching closely: ~20 min·3 must-watch clips
- 0:13 – 4:00Listen
Why put the agent on the canvas
The opening frames what this agent is: it knows your canvas, knows your context, and works right alongside you through two entry points - one in the left rail and one beside your selection. It also notes the agent is in open beta on Pro, Org, and Enterprise, and is heading to FigJam and Slides.
What AI should take over is the part of the work you personally don't like - and that's a different thing for everybody, so the tool can't decide it for you.
This stretch is basically philosophy delivered standing up, with the screen parked on a title slide - listening is enough.▶ Jump to 0:13 - 4:00 – 7:14Listen
The four new things
Announcing web search, MCP connectors (Notion, Slack, GitHub, and writing back to them), composable and shareable custom skills, and shareable chats. Along with a candid product dilemma: an open prompt box makes people assume it can do anything, and one unsupported use case is enough to ruin the whole experience.
Connectors go both ways - after generating designs from a PRD, you can write the changes back to the PRD with references to all the designs.
The capability list is read out loud; the screen only holds bullet text, so there's no need to watch.▶ Jump to 4:00 - 7:14 – 11:00Watch
Two entry points, one real generation
A tour first: a tab in the left rail holding every past conversation, plus an entry point that appears beside your selection when you pick an object on canvas. The plus menu takes images, files, web search, libraries, connectors, and skills. Then, live on stage, a first version of a search results page built from a Notion spec and two reference designs.
You assemble the context yourself - where the spec comes from, which designs to reference, which library to use, all specified before you hit go.
It's all mouse work: what's inside the @mention menu and how the library gets attached only comes across on screen.▶ Jump to 7:14 - 11:00 – 14:15Watch
Parallel prompting and constrained variations
While one task runs, three more go out at once - rewrite the copy with today's live weather, swap the profile pictures, duplicate and flip light/dark - with hand edits continuing on canvas throughout. The other speaker demos Create Variations: don't touch style or fonts, just try layout changes, and add European examples since a lot of customers are in Europe.
Constraining it is more useful than freeing it - only by limiting it to layout do you dare let it run a whole batch.
Several tasks spinning in the left list while the cursor never stops on canvas - that no-queueing feel only comes through in the picture.▶ Jump to 11:00 - 14:15 – 18:30Watch
The output is layers you keep editing
Reviewing the finished runs: the content really did get pulled in, off-system typography and fills are reattached to the closest matching styles, one sentence makes the compass rotate and wobble, and code-backed interactive objects go vertical, become a grid, get a bigger title. Then a collaborator's specific comments are handed to the agent to address in one click.
The deliverable is pure design layers still connected to your design system - it's design layers all the way down, and you can go straight in and change any part of it.
The motion section doesn't land at all without watching - whether the compass wobble, the card reflow, and the bigger title worked is something your eyes judge.▶ Jump to 14:15 - 18:33 – 23:52Watch
How custom skills get written and used
A hands-on look at skills the team wrote themselves, including a feedback skill that bottles a senior designer's review instincts, and Spec Up for filling in documentation. It also shows skills layering: simple skills built on a base skill that captures general best practices. Along the way the speaker deep-links to a specific frame, points out a mistake in the generated result, and sends the agent back to fix it.
A skill is just a markdown file - write your team's conventions as plain-language instructions, no AI knowledge required; share it across the org, or copy one in from another agent in seconds.
The skill file and the result it produces on the design are read side by side, and the jump-to-frame moment where the mistake gets called out is entirely on screen.▶ Jump to 18:33 - 23:55 – 27:15Watch
Review annotations, reuse, and rollout
The feedback skill's final output appears: annotations placed right on the design objects covering typography, hierarchy, and accessibility. The left panel shows a teammate's run still actively going. It closes on reusing skills, how much cheaper exploring gets, and the beta rollout plan.
AI collaboration is built as a team-level capability - chats can be shared so people learn from each other, with keeping one private still one click away.
The annotations are concrete text pinned beside the design; one look tells you whether this review actually yields actionable notes. The availability details at the end are fine to just listen to.▶ Jump to 23:55