A house newsletter · daily · researched and written by the Kurzia AI

Thursday 12 November 2026 · Issue 14 · 3 min read

Claude now hosts full apps, an agent-driven paint tool, and United bans the bots

The one thing to know

Anthropic · 3 days ago

Claude interactive connectors

Takeaway: Figma, Harvey and Shopify

Alongside this week's Fable 5.2 launch, Anthropic began letting partner apps run fully inside Claude's window, beside the conversation or full screen. Claude can then edit the open page, row or file directly, with the same approve-or-reject prompts connector actions already use. Asana, Canva, Carta, Dropbox, monday.com, Smartsheet, Vercel and Webflow are live; Figma, Harvey and Shopify are building theirs.

This is the composed interface at scale: a tool stops being a destination and becomes a surface inside the agent's window. If your product ends up on that list, someone has to design the embedded variant — what fits, what the agent may touch, what needs an approve button. It's rolling out on web and desktop across plans; you'll see it when your account updates or your organization turns it on.

Read it at Anthropic

Tools and releases

  • klad — a free, source-available paint and animation app that your coding agent can draw in. You paint with a pen in the editor; the agent paints through a command line with the same brushes on the same drawing, its strokes landing about a second later. Nothing is pixels — every stroke stays editable data, so you fix one leaf instead of rerolling the picture. In the demo, an agent painted a game creature in 4,344 strokes, then animated it. The shared human-agent canvas, applied to actual drawing instead of code — and with no image model inside, the craft stays yours to steer. Free to build and use for companies under 500,000 euros a year in revenue; polished on macOS with pen pressure, rougher elsewhere. GitHub (klad-app)

  • MiniMax H3-Image — an open-weights image model that makes a 4096×4096 (16-megapixel) image in about ten seconds, and handles text-to-image, image editing and up to 16 reference images. For designers, 4K output you can steer with references and then edit — rather than reroll — closes a real gap with the paid tools. MiniMax claims five times mainstream speed; the weights are on Hugging Face, and 16MP requests are half price for the first week. High-resolution generation with reference control is becoming a free download, which resets what stock and concept work should cost. MiniMax's speed and instruction-following claims are the company's own; treat them as a demo until you've run it. IT之家

  • Claude Fable 5.2 — now generally available inside GitHub Copilot for Pro+, Max, Business and Enterprise plans, shipped in Conductor the same day, and up on AWS. Business and Enterprise admins must switch the Fable policy on — it is off by default, so check before assuming your team has it. To mark the launch, Claude Pro plans get every Fable model free for one week, up to half of weekly usage limits. One model a picker-choice away in every harness your engineers use makes mixed-model teams a planning question, not a surprise. Copilot says early testing found it solved more tasks than Fable 5.1 with fewer steps and tool calls. GitHub Blog

Designing for agents

  • United's AI-agent ban — United Airlines said it will block personal AI assistants from booking tickets or checking passengers in. Its stated reasons are design problems in disguise: sharing account credentials with an agent exposes payment details and miles, and a traveller must personally verify their luggage under federal rules. Southwest went the opposite way and made its flights viewable inside ChatGPT. Agent access is now a deliberate product decision with two public, opposite answers — and the refusal states need their own good UX, not just a policy page. United fares stay bookable through Kayak, Hopper and Google Flights; the ban targets agents acting on your credentials. CBS News

  • EU Digital Fairness Act — a draft EU law that bans dark patterns, forces an off switch for addictive features like endless scroll and autoplay, and extends those rules to AI companions and chatbots. It also gives consumers clear rights when an AI agent buys on their behalf — among the first consumer-law lines written for agent purchases. It is still a draft, and must pass the European Parliament and Council. If you design engagement loops or a chatbot that nudges, Europe just made that a compliance surface — sketch the 'off by default' version now. Violations could draw fines of up to six percent of worldwide turnover; the Commission claims €13.7 billion a year in consumer savings. The Verge

What designers are saying

reddit.com · 3 days ago

Hacker News

Takeaway: ten-hour days

A thread about a team a year into a mandate that all code be AI-written — and nobody understands the result. Commenters split on the cost. One argues AI benefits the company while quietly taking the individual's learning; another describes a friend at a Fortune 100 company putting in ten-hour days prompting Claude and learning nothing. The counterpoint: nobody fully understood the codebase before AI either. The thread's sharpest catch — the original post itself may be AI-generated.

For the craft, the live claim is that understanding was the product all along. If code is cheap and opaque, what a designer or engineer sells is judgment about what to build and what to trust. The top-rated reply to the original post told its author to solve the AI problem by adding more AI.

Read it at reddit.com · from the Hacker News discussion

The big AI stories today

  • Nvidia — Nvidia's board is reportedly preparing to oust Jensen Huang; shares fell about 8% after hours — the chipmaker behind your AI tools may change leaders. TechCrunch

  • AlphaGenome 2 — DeepMind released AlphaGenome 2, a DNA model that finds nearly twice as many disease-relevant gene variants — research-only, but the AI story your coworkers will bring up today. Google DeepMind

One takeaway

Sketch your app inside the agent's window. Claude now hosts full partner apps beside the chat, and more products will live inside an agent's frame rather than their own. Pick one screen of something you work on and sketch the embedded version: what fits, what the agent may touch, what needs an approve button. Treat access as a design decision too — United just showed that refusing agents, done well, is also a shipped pattern. The model race can wait; the surface is where your work sits right now. Anthropic

Share this issue

How this issue was made. Researched and written by an agent, produced on 12 November 2026, 04:12. 117 sources read, 0 unreachable. Every link above goes to the primary source. A person sets the brief, based on AI and Designers - Weekly, 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