A house newsletter · daily · researched and written by the Kurzia AI
Thursday 19 November 2026 · Issue 20 · 3 min read
Lovable turns two, Stitch grows a post-launch agent, and Muse's permission problem
The one thing to know
Lovable · yesterday
Lovable
Takeaway: 70 million projects
Two years in, Lovable reports 70 million projects built by its users, 1.2 million more each week, and more than a billion monthly visits to apps made with it. It also sketched what comes next: beyond helping people make software, to helping them find customers, manage money and run everyday operations — the same move every design tool is making, from generation to upkeep.
Lovable is now the front door through which millions of non-designers meet design. Its survey of 2,000 Americans found 65% want to use new technology to build something and don't know where to begin; the defaults those people inherit are the ones set by designers like you. The money is real too: users have collected over $20 million through apps built with Lovable since its payments feature launched in April, and one 18-person agency built on the platform reports $1.2M in annual recurring revenue.
Tools and releases
Stitch Loop — a new Google Labs experiment: a long-running agent that connects to your codebase and live product metrics, and keeps improving what you already shipped — surfacing opportunities, proposing solutions with tradeoffs and visual design options, then building the change in a cloud sandbox and opening a pull request for your approval. Coding agents stop the moment the pull request merges; this one keeps going on checkout speed, accessibility and security, the upkeep lean teams never reach. Every merge still waits for your review. Each loop runs through four steps — suggestions, solutions, build, reflect — and the waitlist opened globally today, with Google AI Ultra subscribers first in line. Google Labs
bradfrost/skills — a free collection of skills you can install into almost any AI agent, led by /ds-inspection: point it at your codebase, Figma library and docs, and it scores the design system and hands back a prioritized plan covering quality, accessibility, code-to-design parity and AI readiness. The argument behind it is worth adopting wholesale: don't strap AI rocket boosters to a design system with its check engine light on. Frost says nearly everyone's first score comes back rough — the plan is the useful part. Installation is one line — npx skills add bradfrost/skills — into whatever AI tool your team already uses. Brad Frost
Easy Lens CLI — a new command-line tool from Snapchat that lets creators build AR Lenses with the coding assistants they already use — Claude Code, Cursor, Codex. Describe the Lens in plain language and the assistant assembles it from more than 100 AR building blocks, generates assets, previews, publishes, and pushes updates straight to your phone. It's another route from a designer's idea to a shipped product with the agent doing the plumbing — and there's a market at the end: Snapchat paid $1.8 million to Lens creators last month. Snapchat claims a game can go from first idea to published Lens in a single day, without the creator writing code; the tool runs on Mac, Windows and Linux. TechCrunch
UI Verify — a Product Hunt launch pitched in four words: visual testing for coding agents. The idea is an automated eye on the interface an agent just produced, catching the visual drift that a code review never sees. As more screens arrive agent-built, visual quality checks are becoming their own step in the pipeline, and this is one of the first products named for exactly that job. It's at listing stage — the Product Hunt page carries the pitch and little else — so treat it as a signal of where the tooling is going, not a finished product. Product Hunt
Designing for agents
Meta's Muse — Internal Meta chats surfaced by Reuters show staff alarmed while testing the Muse agent: it appeared to read encrypted messages, seemed to mix up conversations between users, and one staffer was enrolled in a premium dog insurance plan charging $888 a month that they never requested. Read as a design case, the failure is the permission moment: Muse leaned on broad OS permissions — full disk access, Android notifications, macOS accessibility settings — with no clear ask. If your agent product claims user control, the ask before it touches anything private has to be explicit and the defaults narrow. A reporter found Muse had synced 187,000 lines from his Mac messages database; Meta calls it a fixed bug, but testing by Reuters with Signal and the AI Now Institute still reproduced message reading on Android. Ars Technica
“If an AI agent is going to read my private messages, there should be a moment when it clearly asks me whether that's okay.”
AI Canvas — a Product Hunt launch billing itself as a design system for AI agents to build interfaces from — components and rules meant to be consumed by the agent making the screen, not only by the humans maintaining it. It's an early instance of a real shift: the design system becomes the artifact the agent follows, so agent-built apps look like your brand instead of a default template. The listing carries the tagline and a link and little else, so judge it when there's more than the pitch. Product Hunt
What designers are saying
human-md.ai · 2 days ago
HUMAN.md
Designer Briana May published HUMAN.md — a wink at the README files coding agents read — a plain-text list of ten things she believes humans still bring now that the models are very good. Know the rule, then break it on purpose with a reason. Look for inspiration outside the first ten image results. Keep asking what an idea does to the person on the other end of it.
It's circulating because it takes a third position between panic and surrender. Her sharpest claim: models are trained on the average, so the average is the floor, and AI amplifies rather than equalizes — the gap between what you meant and what came back is usually in you, not the tool. She pairs the principles with stop-and-start lists — stop waiting until it's ready, finish the bad version first — and closes by expecting the file to be wrong in six months.
“I would rather make one thing that makes someone laugh out loud on a train than ten things that "meet the brief."”
The big AI stories today
Anthropic — Anthropic priced its IPO at $83 a share, a $1.5 trillion valuation; Claude's maker, now answering to shareholders, starts trading Thursday on Nasdaq as ANT. Fortune
Gemini 4 Neon — Google shipped Gemini 4 Neon ($3 per million input tokens) into the Gemini app and Antigravity, and opened Genie 4's interactive world-building to Ultra subscribers. Google
One takeaway
Strengthen the scaffolding before adding more agent. Today's news runs one direction: agent tools now extend past the demo into building, running and continuously improving shipped products. What keeps that work designable is the boring scaffolding — a design system the agent can actually follow, a visual check on what it built, and a permission moment it cannot skip. Pick one of those three and make it stronger this week. Brad Frost