The AI build process gives UX one job — plus Figma Weave, agent payment patterns, and Anthropic's IPO filing

Figma Weave — turns AI image and video generation into connected, tweakable node workflows — and now links to Figma Design: copy a frame into Weave and it stays connected to the original, so brand layouts can guide new work. This turns one-off prompting into a process your team can branch, hand off, and reuse — what design systems did for components, now done for generation. Community publishing is live for packaged Weave tools anyone can run without touching the nodes, and one prompt can run through several AI models side by side for comparison. figma.com

Figma's editing model — is a model Figma post-trained for one job: the small, precise canvas edits designers make all day, output in Figma's own file format. The agent uses it for exact edits and leading third-party models for images and fresh explorations. A smaller purpose-trained model beating frontier models at a focused task is why the agent's edits have been feeling faster and less wrong. Typical agent tasks are about 20% faster than at beta, and designers undo the agent's work 60% less often; an earlier version trained on more than 250,000 edit examples was already competitive with top models. figma.com

Gemini file editing — lets you generate, preview, and edit documents, PDFs, decks, and spreadsheets inside the chat — go full-screen to type your own changes, or highlight a passage to have Gemini rewrite just that. It shortens the draft-and-fix loop for the documents around design work, and shows where every assistant is heading: the chat becomes the editing surface. PDF editing works on web and mobile; every other format is web-only. It's live for all users, in every language where the Gemini app exists. blog.google

Genie 4 — is DeepMind's new world model: type a scene and it renders a place you move through in real time, steered by plain language, with matching ambient sound and dialogue generated on the fly. Near-term it's a pre-visualization tool — film, games, environment concepts you can walk through instead of storyboard — and a preview of what prototypes-as-running-experiences could feel like in consumer tools. Sessions hold together for about two minutes, double Genie 3, and Google admits details and object permanence still drift; it's a research preview with a Project Genie waitlist. deepmind.google

Agent payment approval — shows the person the exact charge an agent wants — merchant, items, total, reason, which card — for an explicit approve or decline before money moves. A timeout counts as a decline, never an approval. The rule set transfers to any irreversible agent action: show the full scope, give approve and decline equal weight, and let silence mean no. Six states run from Requested to Failed; on approval the wallet releases a credential scoped to that one charge, and Stripe Link's approval window expires fast and stays dead. aiuxplayground.com

Virtual card — gives an agent a one-time card locked to one merchant and one amount, so it can pay without ever holding the person's real card number. The credential dies after the charge; a different purchase needs a new one. The wording matters as much as the plumbing: never tell someone the agent 'has their card' when it holds a scoped credential, and never print a full card number in a chat transcript. Visa puts the spend controls on a network token, and Mastercard's Agent Pay has agents pay with signed vouchers inside the preset limit. aiuxplayground.com

Activity log — is the receipt list for agent spending: each row shows date, item, merchant of record, amount, status, and an undo action while the window is open. A purchase lands when money moves, not when the agent merely recommends something. Undo with a stated deadline is the part worth borrowing — it makes autonomous buying feel reversible, and keeps suggestions visually separate from real charges. Five states run Waiting to Closed; shipped examples include Stripe Link, Meta Muse, and Amazon Alexa. aiuxplayground.com

DoorDash's agent numbers — told investors that outside AI agents — Muse, Instinct, Fo — have produced only 'several hundred' orders in total, against roughly 3 million deliveries a day. Worth keeping in your pocket when someone demands an agentic-commerce roadmap: usage is tiny so far, and simple tasks like ordering lunch still happen in the app people already have. CEO Tony Xu says he won't fight gravity but won't be delusional either; Amazon blocks personal agents while Walmart welcomes them. techcrunch.com

“we're seeing maybe several hundred orders in total from some of these third-party agents”

Anthropic — filed its IPO paperwork with US regulators to list on Nasdaq under the symbol ANT, the first pure-play AI company to reach this point — meaning the company behind Claude will soon answer to quarterly earnings. anthropic.com

OpenAI funding — cut its funding target from $30 billion to $20 billion and its valuation aim to $1.25 trillion after tepid investor interest — a sign the money behind the tools in your workflow is getting pickier. trendingtopics.eu

Gemini 4 Carbon — is Google's strongest model, released first to vetted cyber-defender teams before the public — a staged rollout worth noting since models like this reach your product teams within weeks. blog.google

GPT-6.1 Terra — is OpenAI's mid-tier model for routine screen work — looking, clicking, typing — at half the price of its predecessor, a sign that running agents through everyday tasks is being priced like a commodity. thenextweb.com

Claude Code backdoor — A tampered copy of Anthropic's Claude Code extension sat in Cursor's marketplace for nearly 12 hours, planting an SSH backdoor on up to 5,000 machines — the agent toolchain your team auto-installs is a supply chain worth auditing. arstechnica.com