A house newsletter · daily · researched and written by the Kurzia AI
Saturday 14 November 2026 · Issue 16 · 3 min read
ChatGPT's agents can now pay, a stage-by-stage map of shopping UX gaps, and Claude Code goes quiet on agent costs
The one thing to know
TechCrunch · yesterday
ChatGPT Wallet
Takeaway: Face ID or a fingerprint
OpenAI launched a wallet that lets its agents, called dots, shop and finish the purchase for you. Before checkout you see the items, the total and the card, and you approve with Face ID or a fingerprint. It starts with a small group of US ChatGPT Pro subscribers aged 18 and up.
Your card number never reaches the agent or the shop: the charge runs on a one-off payment token issued by the card network. Retailers are split — Amazon blocks agents outright — and OpenAI published a playbook for merchants who want to let them in. Spending limits and pre-approved purchases are promised for later.
Tools and releases
Figma Weave — Figma announced that Weave workflows — repeatable chains of AI generation and editing — can now sit alongside Figma frames on the same canvas. The accompanying essay by Weave's co-founder Itay Schiff hands you the argument for your team: keep an open mode for exploring and a closed mode for committing and automating, and decide deliberately which steps you hand over. Figma's own campaign for Gaussian splat, a 3D tool in Weave, ran on Seedance 2.5, a video model Weave hosts — built after long human riffing on story and style, not one prompt. Figma Blog
“AI is changing the conditions of creative work, bringing enormous possibilities and real losses at the same time.”
Replicate — A free tool on Product Hunt converts an existing Webflow or Framer app into a Claude or Codex project — files a coding agent can open and keep building on. It bridges the design-to-code handoff: the site you already shipped in a no-code tool becomes your agent's starting point, rather than a screenshot it has to guess from. The listing is thin — no pricing tiers, no detailed walkthrough — so treat it as a quick experiment, not a workflow to adopt. Product Hunt
Designing for agents
Agentic Shopping framework — AI UX Playground published a framework that walks a purchase made through an agent in eight stages, from the first search results to the screen where you cut an agent's spending off. Its timing could not suit ChatGPT Wallet better, and its judgement is harsh: the industry has the approve-and-pay moment handled, while the agent choosing the wrong item, order tracking, returns and the spend kill switch are where products still leave shoppers stranded. Each stage names real examples — Alexa shopping, Perplexity — plus a pre-ship checklist. AI UX Playground
Research-in-the-loop — NN/g's new article argues that when agents build features in an afternoon, research shifts from studying users before the build to writing what each release must measure into the spec — the same document your coding agent builds from. The agent then builds the event logging and dashboards alongside the feature itself, so every release produces evidence, while qualitative work explains what the numbers cannot. Author Tony Alicea lists the new skills: spec fluency, directing coding agents, and reading analytics instrumentation. NN/g
subagentStatusLine — A bug report on Claude Code's GitHub says version 2.1.323 removed subagentStatusLine, the only setting that showed each running subagent's live context use, spend, model and reasoning effort, with no replacement and no warning at startup. Agent rows now show little beyond a name and the current task, so anyone running several background agents can no longer tell which one is burning context or money — a reminder that visibility into agent state is a feature people build their work around. Version 2.1.322 still renders the custom row; the reporter asks for the hook back or a replacement surface to draw on. GitHub (anthropics/claude-code)
What designers are saying
bacon roll · 2 days ago
Vibes Don't Maintain
Takeaway: no spec, no test, no document
The author's point: the happy path a demo shows is the smallest part of real software; the rest is seatbelts.
That is the spec argument in story form. When such a prototype fails, there is no spec, no test, no document explaining the intent — and writing that record is design work, not engineering garnish. The author's own team keeps specs and decisions as markdown files in git so a human or an agent can read them back; his friend had a chat history he could no longer find.
“AI does not make you a software engineer if your background is baking bread, any more than your oven at home makes you a chef.”
The big AI stories today
Blackwell fingerprints — A forensic study of published model weights says six new Chinese models carry the fingerprint of export-restricted Nvidia Blackwell chips — more fuel for the export-control fight. Unemployment Labs
LangChain — Agent tooling keeps pulling money: LangChain raised $300M at a $3B valuation for LangSmith, and agent-security firm WhiteMatter and sales-agent Rox also announced rounds. LangChain
FAA SMART — The FAA extended SMART, its AI route-advice system for air traffic controllers, to all 22 US centres; the controllers' union was never consulted. Ars Technica
Claude status — Anthropic resolved two Nov 13 incidents that raised errors on claude.ai, the Claude API and Claude Code; some scheduled tasks ran over two hours late. Claude Status
One takeaway
Design the off-switch before the checkout. Agentic spending just went mainstream, and the settled part is only the approval tap — the wrong purchase, the order trail and the revoke screen are open territory no pattern owns yet. Sketch the revoke state before you polish the buy button: it is the screen people will hunt for the day an agent misfires, and whoever gets it right sets the convention everyone else copies. AI UX Playground