Figma puts real code on the canvas — plus Anthropic's junior-hiring data and fresh patterns for agents

Code layers — Figma's open beta for prototypes that actually run on code, explored together on the shared canvas. Bring in a codebase from GitHub or a local folder, or capture a live page with the Chrome extension, then duplicate versions side by side and compare them as a team. Design-to-code stops being a one-way handoff: you refine interactions by prompting the Figma agent, pull screens back into editable layers with Extract designs, and push changes to a code branch for review. Merkle reports 30% faster iteration and 50% less spec documentation; creating and editing is free on paid plans through November 4. Figma Blog

Figma agent — Now generally available on the canvas, with steadier instruction-following and longer-running tasks. Library owners can attach guidelines — say, red is reserved for destructive actions, never decoration — so agent output follows house rules automatically. The shipped use cases are the story: design-system documentation, meeting-feedback capture, and motion that non-specialists can produce on brand. Uber's team publishes component documentation a single designer finishes in an afternoon — work that took several people months; Granola's designer has the agent annotate the canvas with feedback pulled from meeting notes. Figma Blog

Payload — The open-source CMS whose team joined Figma is now inside Figma Make: prompt an app and it sets up a database, file storage, user login and an admin panel, all hosted by Figma. Marketers edit content through the admin panel instead of prompting or filing tickets, and content stays bound to your design system — designers keep the brand, non-designers keep the keys. Available on all plans over the coming weeks; the hosting layer is free while in beta. Figma Blog

Notion 3.8 — A release built around durable context for agents. MCP connections let outside agents read the workspace and update docs and trackers, Notion's own agents draw on 150+ tools, and a Skills API syncs skills to ChatGPT, Claude and GitHub. Upgraded version history now attributes every change to the person or agent who made it — the governance layer shared-agent workflows have been missing. On Business and Enterprise plans, an existing ChatGPT Plus or Pro subscription can cover AI use in Notion at no markup. Notion

AI adoption rituals — A Figma piece on four rituals that spread AI fluency beyond the lone power user: hour-long weekly study groups at Figma, a biweekly show-and-tell at Sephora, cross-training between design and engineering at Mercari, and prototype-first critiques at Atlassian. If two people on your team hold all the AI skill, these are the formats that measurably moved the baseline at large companies. Atlassian's rule: every crit opens with a clickable prototype, explained only afterward; Sephora's forum requires presenters to share what failed, not just the polished result. Figma Blog

Krea 3 — An AI image model whose whole pitch is turnaround: 2K edits in under 8 seconds, per its Product Hunt launch. For moodboards and visual exploration, edit latency decides whether you iterate or settle — worth testing against your current tool. It's a Product Hunt listing, so treat the speed claim as marketing until you've run it. Product Hunt

Claude Opus 4.5 in Lovable — Lovable says Claude Opus 4.5 now powers planning in chat mode and core code generation, at no additional cost to users. Its most successful builders map builds out in conversation before editing — better planning is where design intent gets set; Lovable claims 15% fewer edit mistakes. Lovable's blog already lists newer model announcements, so check what's current before comparing notes. Lovable

Siri AI — A hands-on NN/g test argues a useful assistant needs personal context and the capability to use it — plus an interface that shows its limits. Apple's new Siri AI found trip photos and texted restaurant tips that ChatGPT had no route to. The failures are the lessons: it missed a message because it can't read Signal, then said it found nothing, without naming which apps it had searched. It even confirmed the right destination, then opened Maps from the wrong starting point anyway — test agent tasks end-to-end, including the destination app. Nielsen Norman Group

Context libraries — An NN/g study of expert Claude users finds them curating "context libraries" — markdown files of procedures and standards that agents read and edit. Almost nobody edits the files by hand; they ask the agent to do it, then judge the results. The pain is the finding: files go stale, go missing, and external connections break, because no product helps users curate context. That's an open design problem for anyone building agent tools. The most elaborate library in the study was engineered to load as little as possible, with CLAUDE.md files acting as index cards describing each folder. Nielsen Norman Group

Muse — Design director Alex Cornell walks through Muse's design on the How I AI podcast. Every activity renders three ways — an emoji for casual users, a task list for the detail-minded, raw logs for engineers — and the main chat is one ongoing relationship, not separate threads. Generative UI works here because the structure stays fixed while the content personalizes. His test for every decision is portable to any agent project: ask how it would work between two humans. He pointed Muse at the Instagram posts he'd bookmarked over the years and had them organized into a shopping guide styled after an old in-flight catalog. How I AI (Lenny's Podcast Network)

“How would this work if you were talking to another human?”

Synthetic users — NN/g ran 120 AI-simulated survey respondents against 604 humans on the same open-ended question. Both reached the same headline conclusion, but the synthetic answers collapsed onto roughly a dozen terms where humans used more than fifty. If you're tempted to swap real participants for simulated ones, note the flattening: 93% of synthetic product managers used the same terminology family, versus 44% of the real ones. NN/g's verdict: synthetic-user research is not user research — treat its findings as hypotheses to test, never as evidence. Nielsen Norman Group

“Treat synthetic findings as a starting point rather than as evidence of what users think, say, or do.”