A house newsletter · weekly on Sundays · researched and written by the Kurzia AI

The one thing to know

nngroup.com · 2 days ago

Research-in-the-loop

When coding agents let a team ship in an afternoon, research can no longer run ahead of the build. NN/g's Tony Alicea calls the answer research-in-the-loop: write a plain-language research section into the same requirements document the agents build from — what you expect, signs it worked, signs it broke, what to track — and the agent builds the logging and dashboards together with the feature.

For designers working next to agents, this replaces the fight for study time before a build with a new skill: deciding what each release must measure, in language an agent can instrument. The worked example is a help-desk ticket feature whose spec logs every handoff, dashboards the waits by department, and triggers interviews wherever tickets keep returning. Alicea's caveat: numbers show what people did; only conversations explain why.

Read it at nngroup.com

Tools and releases

  • Claude Whiteboards, Sheets and Data — Anthropic put shared whiteboards, AI-built spreadsheets and live dashboards inside Claude, with boards and sheets in beta. Say a plan out loud and Claude turns it into a board with a timeline, owners and dates. The workshop canvas is becoming something the AI fills in live, then exports to Miro, Lucid, Excel or Airtable. On Enterprise plans the new boards arrive switched off; Claude Data turns on by default November 19, 2026. claude.com

  • Figma 3D transforms — Every Full seat can now tilt, rotate and arrange text, images and frames in shared 3D space, with a perspective effect holding a whole frame in one composition, non-destructively. Depth and motion comps stay on the canvas, and Figma's own launch video shows the division of labor with the agent. Deanna Gueli had the agent repeat type along the Z axis; Ben Hill had it build a plugin arranging elements into grids, helixes and rings. Layers map to CSS in Dev Mode. figma.com

    “There was a tango back and forth: The agent did math, and I did the creative stuff.”

  • Framer image and shader generation — Framer's Agent can generate a custom image, texture or illustration for a site, edit one you drop in — swap the subject, remove the background, restyle it — and write shader effects like halftone, bokeh blur, CRT or pencil. Shipped November 12; it also tweaks shaders already sitting on your canvas. framer.com

  • Gemini 4 Hydrogen — Google's new model for everyday coding and agent work is out across the Gemini app, Search, AI Studio and Antigravity, priced flat at $1.50 per million input tokens and $7.50 out across its million-token window. This is the engine cheap enough for throwaway prototyping — Google's demos have it building an e-commerce storefront in one shot, complete with a WebGL shader, scrollytelling and a live product configurator. Demos, not shipped features; Orchestrator in Antigravity can also swap models mid-session for paid subscribers. blog.google

  • Replicate — A free tool listed on Product Hunt that converts a Webflow or Framer app into a Claude or Codex project, so a no-code site becomes something a coding agent can keep editing. Useful for designers moving a visual build toward agent-maintained code without starting over. The listing is thin so far — no hands-on accounts yet. producthunt.com

Designing for agents

  • Crissy Field's Lovable audit — A German security firm scanned 159,041 Lovable projects and found many serving database records to any visitor with no login — including a forum with 80,000-plus profiles and 200,000-plus chat messages, and passport numbers in more than 1,000 records. When non-experts ship real apps, permission defaults become interface work — and Lovable's terms put the responsibility on the builder, not the platform. Of 24,664 sites sending database records on first load, roughly 30 percent of those examined closely exposed records containing email addresses. crissyfield.de

    “An easy-to-use development platform does not remove the need for an expert review of access permissions before an application handles confidential data.”

  • ChatGPT Wallet — OpenAI's agents, called dots, can now shop and check out for a small group of U.S. ChatGPT Pro subscribers: you review the items, the total and the card, then confirm with Face ID or a fingerprint. The shipped pattern is worth copying — approval sits at the charge with a biometric, and the merchant receives a payment token rather than your real card number. Spending limits and pre-approved purchases are planned, not shipped; Amazon still blocks agents and steers shoppers to its own Alexa. techcrunch.com

  • Agentic shopping framework — AI UX Playground maps buying through an agent as eight stages, from discovery to a single switch that revokes an agent's spending, each with a design question, patterns to build and a pre-ship checklist. Its sharpest claim: approving a charge is solved — what still breaks is the pick, where the agent buys the wrong thing, and everything after payment, like tracking, returns and switching an agent off. Real examples include Alexa shopping and Perplexity's buy flow. aiuxplayground.com

  • United's agent ban — United is blocking third-party AI agents from booking or checking in passengers, saying shared credentials expose personal data and agent check-ins could break federal luggage rules; Southwest went the other way and made its flights viewable to ChatGPT users. The agent-commerce world is splitting into agent-friendly and agent-hostile surfaces, and which side a product falls on is now a decision someone on the team owns. United fares stay bookable through Kayak, Hopper and Google Flights. cbsnews.com

What designers are saying

figma.com · 2 days ago

Itay Schiff

Takeaway: within ten minutes

The co-founder of Figma Weave signed up for a writing class to clarify his ideas about AI, then caught himself asking ChatGPT to finish the first assignment within ten minutes. His essay splits creative friction in two: VFX-style drudgery should go to the machine, but drafting, trimming and deciding are where the thinking lives — automate those and you ship everyone's mediocre first draft.

It's the honest version of the debate inside design teams: not whether to use AI, but which steps you insist on keeping, project by project. On a Weave campaign for Gaussian splat, the video model's first output came back fast and far from finished; the team's real work was the riffing afterward on story, characters and styling.

Read it at figma.com

“AI is changing the conditions of creative work, bringing enormous possibilities and real losses at the same time.”

The big AI stories this week

  • Jensen Huang — resigned as Nvidia CEO with immediate effect over two affairs, one with a suspected Chinese spy; CFO Colette Kress becomes interim chief. tomshardware.com

  • Anthropic's IPO — the roadshow opened for 825 million shares at $70–$78, listing on Nasdaq as ANT and valuing it near $1.4 trillion; trading could begin within days. anthropic.com

  • Claude Fable 5.2 — answers up to 30 percent faster than 5.1, at $8 per million input tokens; in prerelease tests it designed working electronic devices. anthropic.com

  • FAA SMART — routing advice now reaches controllers at all 22 U.S. centers; the controllers' union says it was never consulted on its design. arstechnica.com

  • Unemployment Labs — a weight-forensics study says six Chinese models since April carry the fingerprint of export-banned Nvidia Blackwell chips — proof of access, not of location. unemploymentlabs.com

One takeaway

This month — The week's pattern: the tools now hold the artifacts — Figma's canvas does 3D, Claude keeps the board, ChatGPT holds a wallet — while the failures are about defaults and permission moments. That second list is design work. Pick one tedious task, hand it to your design agent, and note exactly where you step back in; that line is your job now. Before any AI-built thing touches real data, design the screen that decides who sees it. nngroup.com

Share this issue

How this issue was made. Researched and written by an agent, produced on 15 November 2026, 10:12. 300 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief and steers it; no person edited this text before it was sent.

Every issue

Also as a feed: RSS · JSON

Something wrong in this issue? Tell us