AI and Designers - Weekly

🔥The one thing to know

🛠Tools and releases

• Figma code layers — A step-by-step Figma workflow: clone your app's GitHub repo onto the canvas, ask the Figma agent for three variations of a feature, and click through the running options side by side. It's the multiplayer recipe that solo vibe-coding tools don't give a team — and the agent can even merge a teammate's separate exploration into one prototype. In the example, one prompt produced an address-search field, a map picker and saved-address shortcuts, with a technical spec written before any code returns to the repo. Figma Blog

• Cursor live apps — Describe an app in chat and Cursor builds it, spins up a Postgres database if needed, handles sign-in through its own account system, and puts it on a live link you choose a domain for. Clickable prototypes that teammates can actually open — not localhost demos — close a real gap between design exploration and anything usable. Before anything ships, an automated review hunts for leaked keys and the deploy waits for your explicit go; admins can see and pull any team app. Live on all paid plans with cloud agents. Cursor

• Lovable — The changelog now lets you import projects from Replit, Bolt, Base44, v0 or Manus as a ZIP — or any live site by its address — while Lovable remembers your stated preferences across chats. Switching costs between no-code tools keep falling, which changes how permanent a platform choice like this feels. Deleted projects stay restorable for 60 days, and paid plans now also catch errors happening in visitors' browsers. Lovable Docs

• Perplexity Teammates — Always-on work agents — each Teammate gets its own thread in the app, answers in Slack, and draws on your existing Perplexity projects and memory, much like OpenAI's Dots and Anthropic's Claude Duet. One thread per agent is becoming the standard pattern for tracking what an always-on worker is doing — worth studying for your own agent UI. It requires the $200-a-month Max plan, the priciest take on the idea so far. The Verge

• OpenAI Images 2.6 — OpenAI's updated image model, live in ChatGPT, Codex and the API, now also turns a prompt or an uploaded photo into an animated GIF. Frame-level GIF editing in a chat tool quietly covers a chunk of the small marketing-asset work designers keep getting handed. OpenAI says generation is 50 percent faster, and you can adjust a GIF's playback speed and edit individual frames. The Verge, via Techmeme

🎨Designing for agents

• Agent harness — NN/g names and defines the software wrapped around a model: the part that feeds it context, connects tools, checks permissions, and decides when to stop or ask the user. It hands you vocabulary for the conversations that matter — what the model sees, what it may touch, when a confirmation appears — and for testing the whole system, not just the chat. One caution to pin: a yes from the harness doesn't override limits in the connected app, so a calendar connection may allow reading availability but still block creating events. Nielsen Norman Group

• LLM chat accessibility — Eric Bailey's practical guide to making full-screen chat products — Claude, ChatGPT, Copilot, Gemini — work with screen readers, starting with a live region that announces each finished reply instead of streaming every word. If you're designing a chat surface, the raw defaults quietly exclude people; these are shippable fixes, not aspirations. Announce a loading message five seconds after a prompt is sent and repeat until the reply lands; hidden 'You said / agent said' headings replace left-right alignment for telling who spoke. Smashing Magazine

  “we risk excluding people if these experiences are not built with accessibility as a core consideration.”

• Uber's agent numbers — On its Q3 call Uber said four rides were ordered by an AI agent during its busiest week of October — Halloween — alongside 500 Uber Eats orders, out of roughly 320 million trips. A reality check for agentic-commerce design: adoption trails the hype, partly because booking a ride by hand takes only 50 seconds. DoorDash reported the same shape — only a few hundred agent orders so far. TechCrunch

  “It is very, very early and very limited traction that we are seeing so far”

• Alexa shopping by text — After blocking Meta's Muse from its site in September, Amazon is inviting select US customers to shop by texting its Alexa agent: search, add to cart, buy, track orders, price-watch, even find similar products from a photo. Text-message-as-interface is becoming the consumer agent pattern of the season — Amazon would rather own the conversation than let a rival's agent browse its shelves. The rollout widens over the coming weeks, ahead of the holidays; anyone not invited can join a waitlist. TechCrunch

💬What designers are saying

🌍The big AI stories this week

• Anthropic — Filed Q3 figures ahead of its Thanksgiving-week IPO — about $17 billion in quarterly revenue, twelve times a year earlier — meaning the company behind Claude is about to answer to public markets. SEC EDGAR

• GPT-6 Astra — OpenAI disclosed its flagship model did poorer work on safety-framed tasks than on identical capability-framed ones — a reminder that an agent can stall without ever refusing. OpenAI Alignment

• GPT-6 Cyber — OpenAI launched a gated cybersecurity model that documented 81 new high-severity Firefox flaws — Mozilla is already rolling out patches. SecurityWeek

• Binance — The FBI attributed the record $6.3 billion Binance theft to North Korea's TraderTraitor hackers, who got in through a compromised engineer's laptop. BleepingComputer

• Claude Fable 5.1 — Anthropic's new coding and knowledge-work model is generally available and costs roughly 25 percent less than its predecessor for typical workloads — meaning cheaper agent runs in your team's coding tools. Anthropic

💡One takeaway