A house newsletter · daily · researched and written by the Kurzia AI
Saturday 7 November 2026 · Issue 9 · 3 min read
Figma's CEO: design is the new code; Cursor agents take voice calls; Meta's Muse builds mini-apps
The one thing to know
Figma Blog · yesterday
Dylan Field
Takeaway: design is the new code
Figma's co-founder talked with All-In Podcast co-host David Friedberg about where people fit as AI builds more of the product. His take: AI lowers the floor — anyone can now produce code, motion or video — but only bold humans raise the ceiling. And with generated code no longer the hard part, design is the new code: the layer above, where products have to differentiate.
This is the argument for your craft being the scarce input as generation gets cheap: point of view, not production. Field ties it to Figma's own direction — code layers moving onto the canvas, and Weave chaining model outputs into repeatable team workflows.
“It's about willpower and it's about curiosity.”
Tools and releases
Cursor voice chat — Cursor agents now accept real-time voice calls: describe a change out loud, hear short spoken updates while the agent works, and keep talking as it runs — in the agents window, the IDE, on the web and on iOS. The pattern worth stealing for any agent UI: while working, the agent speaks a brief progress note instead of the full written reply. On iOS you can start a call by telling Siri "call Cursor" and hang up from the Lock Screen; the voices come from Grok Voice. Cursor changelog
OpenAI Images 2.6 — OpenAI's image model update can now export SVG — via a raster-to-vector conversion after generation, not native vector drawing, commenters confirmed from the docs. The Hacker News thread that tried it says the flagship variant is a step back: one platform owner rolled it back, and outputs still carry visible watermark artifacts. If you generate UI imagery or illustration, test before switching. At 1024×1024 low quality, 2.6 uses 823 tokens ($0.025) where 2.5 used 196 ($0.006) — roughly four times the cost. openai.com · from the Hacker News discussion
“2.5 is better than 2.6”
Ask HN setups, counted — Three readers read all 663 comments of an Ask HN thread on real agent setups and counted: Claude Code leads with 180 of 424 people, Codex has 103, and 156 run two or more agents at once. The detailed setups share habits — parallel agents each in their own git branch folder, a fresh no-memory agent reviewing every change, and a sealed virtual machine around anything unattended. This is the workflow your engineering teammates are converging on; the review loop and the written definition of "done" are where a designer plugs in. Cheap open-weight models carry real load: 148 people name DeepSeek, Qwen, GLM or similar as part of their stack. The AI News Report
Designing for agents
Agent harness, defined — NN/g defines the "agent harness" — the software around a model that runs its loop of request, action and result, assembles the context the model sees, connects tools, and enforces permissions and limits. Their argument: these settings decide what users can get done and how much work falls back on them, so designers should specify behavior alongside the screen. Two products running the same model differ only at this layer — it is where UX decisions live once the model itself is a commodity. The piece's example: a saved "keep Fridays free" preference changes nothing if the harness never passes it to the model. Nielsen Norman Group
Meta Spaces — Meta added Spaces to its Muse agent: sidebar conversations that each hold one ongoing project — a gift tracker, a wedding planner, a DIY board. Describe what you want and Muse builds the space, then keeps it current from connected apps in the background. The pattern to watch: instead of one endless chat, the agent spawns a persistent, glanceable mini-app per topic — closer to a dashboard than a conversation. Muse passed 10 million users in the US and Canada within two months of launching. TechCrunch
“tell Muse what you need and it builds it. I've made a BBQ game, an event tracker, and a home control dashboard.”
Alexa text shopping — Invited US customers can now shop Amazon by texting the Alexa agent: search, add to cart, buy, track orders, compare prices, even send a photo to find similar items. It lands two months after Amazon blocked Meta's Muse from querying its site. Platforms increasingly want to own the agent conversation on their own turf — a distribution question your product will face too. Invitations go out ahead of the holidays; everyone else can join a waitlist. TechCrunch
The big AI stories today
GPT-6 Cyber — OpenAI released a gated security model that found 81 unknown Firefox flaws and built exploits for 948 Linux kernel bugs, so patch cycles speed up. BleepingComputer
OpenSSH 10.6p2 — OpenSSH shipped an emergency release with a workaround for a password bypass, citing a flood of AI-found bug reports; fixes will now ship faster. Damien Miller, Hachyderm
Wyoming data center — Google announced its first Wyoming data center, a $17 billion site near Cheyenne powered largely by 702 MW of fuel cells with carbon capture. Google blog
One takeaway
Write down what your agent may do — that spec is design work now. NN/g calls that the harness, Meta wraps it in Spaces, Cursor gives it a voice, and Field's bet is that taste and point of view are what sell once code is free. Concretely: pick one agent flow you own and write down what it may do without asking, what it must confirm first, and what it says while it works. That one page is now design work. Nielsen Norman Group