A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 4 November 2026 · Issue 6 · 4 min read
Canva cuts 22% of staff; Siri AI opens to every developer; Netflix ships a mood-based Discover tab
The one thing to know
Canva cuts 22% of staff to 'refound' itself. Canva is letting 1,271 people go — 22 percent of its staff — and folding its product, growth and platform teams into a single product organization, CEO Melanie Perkins announced yesterday. Many remaining roles change or move to central hubs, with Sydney becoming the main home for product. From January, most hybrid staff are expected in an office two to three days a week.
It lands a day after Squarespace cut 19 percent, and the same force is behind both: AI can now produce much of what template tools used to sell. The friendliest design-tool company is reorganizing into smaller, more autonomous teams around that reality. The Verge
“Sadly, it will also mean saying goodbye to teammates and friends who have helped build Canva into what it is today.”
Tools and releases
Siri AI — Apple's new assistant now accepts integrations from any developer, in a preview inside the iOS, macOS and visionOS 27.3 betas. Apps plug their content and actions in through App Intents, and a new 'skills' idea chains several actions into one request — check a delivery, book a table, place an order. Anything an app surfaces can appear in Spotlight's semantic index, so users find it by asking in plain words. For designers, this is another surface where your screens get bypassed by a conversation — the deliverable becomes the actions your app exposes, not the layout around them. App Store submissions open in the spring; it's unavailable in the EU for now. Apple Developer News
A measured answer to 'which framework is best for agents'. Nolan Lawson had the same agent build the same small Pokémon app in twelve JavaScript frameworks and priced each build. The cheapest runs came from the smaller, server-first setups — Hono+htmx at $2.67 — while Angular cost $4.97 and failed to produce a working app twice. Astro's output shipped zero JavaScript; Next.js sent 142 KiB. When agents write the frontend, framework choice stops being a matter of taste and becomes a cost-and-payload decision you can actually measure and argue with evidence. He gave every framework its official agent tooling to keep the comparison fair. Nolan Lawson
Scrawl — a newly launched tool for the moment a bug is easier to show than to explain: you capture what you're seeing as a marked visual frame and pass that frame to your coding agent as context, instead of typing a paragraph about a layout break or a misplaced element. Most agent workflows are still text-in, text-out. Anything that lets a designer drop an annotated image next to a prompt shortens the loop between spotting a problem and getting it fixed. Listed on Product Hunt on Monday. Product Hunt
Designing for agents
Netflix AI Discover — a new tab, rolling out to US English users on supported TVs, that suggests things to watch by mood rather than title. You pick a theme — 'something light', 'scare me', 'family movie night' — then narrow with more specific ones until something lands. Each title carries a 'Tune Your Recommendations' button beside Play, to save what you like or rule out what you don't. That per-item button is the pattern to steal: cheap, reversible feedback exactly where the decision happens, not buried in settings. Netflix openly calls the tab a test kitchen where feedback decides what graduates to the main app. Look for a red gem icon labelled 'New' in the TV navigation bar. TechCrunch
Falcon Guardian — CrowdStrike launched a kill switch for company AI agents that misbehave, plus an integration with Claude Cowork — the premise being that security software watches what internal agents do and can shut one down mid-task, per CNBC. Expect clients to ask where yours is. Announced Tuesday. Techmeme
Claude Code 2.1.317 — the latest Claude Code release adds two small consent moments worth copying: a warning when a scheduled cloud task quietly picks up data connectors you never asked for, with instructions for removing them, and a prompt before a tool call sends your account email address out to the network. Autonomy products keep drifting toward act-first, mention-it-later. These flag what the agent granted itself at the instant it happens — the pattern your own agent UI should follow. Shipped November 3; the release also improves /design-sync, which now reports up front when a design system has moved. GitHub (Anthropic)
What designers are saying
An engineer's honest year with agents: magic, fatigue, taste. Ben Stolovitz, now at Salesforce AI Research, has updated his annual account of working with AI. His workday is supervising five or six coding agents, each in its own copy of the code, at about $3,000 a month in tokens. His verdict cuts both ways: agents prototype features in twenty minutes and clear away drudgery, but left unsupervised they build over-architected slop — and juggling six robot conversations is exhausting in a way coding never was.
The craft note matters most here: his most satisfying work this year was hand-replacing roughly 4,000 lines of agent-written code with a few hundred of his own, because his version was easier to test and generalize. He also drafted a personal AI policy to ground his daily use, and still refuses to let AI write his prose. He reviews agent pull requests by architecture and recurring patterns now, not line by line. Ben Stolovitz
“[AI writing] is as good as it was 3 years ago. Which, when you consider how much everything else has changed since then…”
The big AI stories today
ChatGPT — has 1.2 billion weekly users and answers about a fifth of Google's search volume, OpenAI says — the window most people will meet your AI feature in. bigtechnology.com
Salesforce and Cresta — signed a definitive agreement for Salesforce to buy voice-agent maker Cresta for a reported $2.7 billion. Phone agents for brands like Marriott and Comcast are going mainstream inside Service Cloud — voice is the next agent channel to professionalize. Closes in Q1 of fiscal 2028, pending regulatory approval. Salesforce
Google in schools — will switch its AI tools off by default for K-12 students next year — but schools that keep AI off also lose Google Search. Forced bundling presented as choice: a consent design with real teeth, worth studying for how your own opt-outs read. Education VP Chris Phillips: 'Search and AI are deeply integrated.' TechCrunch
OpenAI Codex — Hacker News users report the setting that feeds Codex chats into model training appearing switched back on without consent; OpenAI has not confirmed. If client or company work goes into Codex, it's worth checking the data-sharing toggle in ChatGPT settings today. The thread is a report, not a verified finding. Hacker News
One takeaway
Sketch the stop button before anyone asks for it. Today's agent news shares one theme: products are racing to give agents more autonomy while quietly building the moments that limit it — a warning when an agent grants itself connections it was never given, a per-item 'never suggest this again' button, an enterprise kill switch sold as a feature. The approval-and-revocation screen is becoming a core deliverable, not a security afterthought.
For whatever agent feature you ship next, sketch three states first: what the agent is about to do, what it did without asking, and how the user stops it for good. If you can't draw the third state, the agent has too much autonomy. Techmeme