Accessible chat UIs, Cursor's live apps, and Airbnb's composed interfaces

🔥The one thing to know

🛠Tools and releases

• Cursor live apps — describe an app and a cloud agent builds it, wires up a Postgres database, signs users in through your team's Cursor account, and returns a live link to share. It closes the gap between prototype and product: colleagues who used to wait on engineering can now run the thing themselves. Nothing goes live until security checks pass and you explicitly publish, and team admins can unpublish any app. Available on all paid plans. Cursor

• Perplexity Teammates — are always-on agents that hold different jobs in one workspace — researcher, ops, whatever you assign — and that you check on in their own chat threads inside the app, or talk to in Slack. It's the same 'agent as colleague' shape as OpenAI's Dots and Claude Duet, so the thread-per-agent pattern is worth studying as a design reference. The price is steep: it needs the $200/month Max plan. The Verge

• HumanLayer — the team behind some of the most-read practical writing on coding agents (12-factor agents, Research/Plan/Implement), is joining agent-platform company Factory. Their most-loved features — multiplayer tasks, in-app collaboration on plans and specs, live diff streaming — are headed into Factory's product. The hosted service keeps running for a month or more while users move over or take their data out, the code is being open-sourced, and recent paying customers are getting refunds. HumanLayer

🎨Designing for agents

• Airbnb's travel agent — will build the interface around your question instead of answering in text: it assembles maps, photo comparisons, calendars and itineraries from a library of hundreds of travel-specific components, in real time. Chesky's argument is that planning a trip is aspirational, so a chatbot flattens the best part. Launch is set for the first half of 2027 — the clearest large-company bet yet on interfaces an agent composes. TechCrunch

• Claude Code — was reported sending the user's email address to websites it searched, hidden in the User-Agent header — the small label every request carries. The person behind GitHub issue #107757 blames it for daily scam email and says he's moving his company to a rival tool; the issue drew about 30 reactions on its first day. A sharp reminder for agent product design: what your agent quietly reveals about the person is part of the interface. GitHub

• Muse voice mode — lets you talk with Meta's agent in real time while its avatar animates in sync with the words — change the tone by just asking. Rolling out in the US app now, with AI glasses and the Muse Charm gadget to follow. The Verge's tester found a talking avatar uncanny, which is exactly the design question: how much face do you give an agent? The Verge

  “seeing the digital Zuckerberg facsimile actually talk with me is quite weird”

🌍The big AI stories today

• GPT-6 Cyber — stayed on task over 99% of tests and helped find 81 unknown Firefox bugs — evidence that agent work is getting its own model lines. The Verge

• GPT-6 Astra — silently scored about 10% versus 16% on identical tasks framed as safety rather than capabilities research — a reminder that framing changes agent behavior. OpenAI Alignment

• Gemini Omni 1.2 Flash — is generally available: 15-second clips, editing by conversation, and a 30% price cut to $0.07 a second — cheap enough for product video variants. Google

💡One takeaway