A house newsletter · daily · researched and written by the Kurzia AI
Friday 20 November 2026 · Issue 21 · 3 min read
Onlook's canvas for real code, Gemini 4 Neon in Copilot, and Anthropic's record IPO
The one thing to know
GitHub · yesterday
Onlook
Takeaway: 27.4k stars
A free desktop app from a Y Combinator-backed team that puts your actual running product on an infinite canvas: you shape the interface visually while an AI agent edits the code underneath, and Git quietly tracks every change. Onlook ships no model of its own — you plug in the Claude Code or ChatGPT/Codex plan you already pay for.
It is the first serious open-source answer to Figma-plus-an-agent, and it removes the export handoff entirely: what you drag on the canvas is the production code, not a picture of it someone has to rebuild. The repo has 27.4k stars; the app opens any Git repository, runs your own dev server, works on a safe copy so your folder is never touched, and currently ships for macOS only.
Tools and releases
Gemini 4 Neon — is now generally available in GitHub Copilot on Pro+, Max, Business and Enterprise plans, built for agent-style coding. Google says it answers sooner and makes fewer unrequested edits than Gemini 4 Carbon, at a lower cost per task. If your engineers build from your designs with Copilot, this is the default engine arriving under that work — notice it when output speed or quality shifts. It bills at Google's list price — $3 per million tokens in, $15 out — and arrives switched on for every eligible plan. GitHub Changelog
Lovable — now plugs into Adyen and Airwallex, joining PayPal from earlier this month, so a generated app can take cards, local methods like iDEAL, subscriptions and refunds without hand-written code. Checkout is where a Lovable-built app either earns trust or loses it; these connectors give the payment screens you design real rails to attach to. You add one from Connectors or just ask Lovable inside the project. Lovable Documentation
DeepSec — is v0's shipped security review: ask in chat and it scans your changes — or the 50 riskiest files in a project — for missing login checks, injection and unsafe handling of user input. The pattern is the prize: every finding is re-checked against the whole codebase, so an alarm gets dismissed when another file already handles it. That is how agent error states should behave. A review stops after 15 minutes, spends credits, and edits nothing unless you approve. v0 Docs
Designing for agents
Google Search — The web's most minimal page now has a slim strip on the left that unfolds into your AI Mode past: earlier threads listed, resumable and searchable, with a hook into Gemini Notebook. Google is treating an AI conversation as a document people return to, not a one-shot answer — the persistence habit your own product's chat threads will be measured against. Product lead Robby Stein says tests showed the rail helped people make progress and keep exploring; it is reaching desktop users worldwide. The Verge
“people are coming to Search with more questions that aren't just one and done”
Tecla — launched on Product Hunt as a proactive agent with no chat window: the todo list is the product, tasks appearing and getting ticked off as the agent works. A counter-pattern worth watching: ambient agents may suit an auditable list better than a transcript, and a list is the quietest trust surface an agent can have. The listing says little about how it chooses its tasks, so read it as a pattern signal, not a finished product. Product Hunt
What designers are saying
robertpeake.com · yesterday
'Verifier's law' predicts what AI takes next — and it isn't the 'human' stuff
Robert Peake, a poet who spent 25 years in software, argues AI automates work that is easy to grade and easy to agree is valuable — which is why code fell before poetry, and why commercial illustration is falling now.
The sting for designers: specs, mockups and anything with clear acceptance criteria are gradable, so agents will eat them; outcomes — adoption, buy-in, judgment — are not, and that is where the role moves. He cloned his own voice for about $5 and his wife of nearly 25 years couldn't tell it from him; among Society of Authors members polled, roughly one illustrator in four and more than one translator in three report already losing paid work to generative AI.
“AI researcher Jason Wei coined "verifier's law": the easier a task is to check, the easier it is to train AI to do it.”
The big AI stories today
Anthropic — opened at $112 on Nasdaq, 35% above its $83 offer price — the biggest IPO ever; the company behind Claude now answers to public shareholders. Trending Topics
A job ad shows Apple building an agent that runs your devices. is hiring to build an agent that codes, browses, controls apps and coordinates subagents, per Bloomberg — Siri may finally act, not just answer. Bloomberg via Techmeme
Gemini — can now drive apps, files and the web on your Mac in the background (US Plus, Pro and Ultra subscribers); the screen-acting agent pattern keeps spreading. TechCrunch via Techmeme
One takeaway
Point a canvas at your real repo this week. Today's loudest launch, Onlook, bets designers want to shape the real product rather than hand over another throwaway mockup — and it is free, so testing that bet costs an afternoon with the agent plan you already pay for. And as you brief AI-built work, remember what falls to machines: whatever has checkable acceptance criteria. Spend your judgment where checking can't reach. GitHub