Each story is 3 lines: what happened, and why it matters to someone designing an AI product. Picked every morning and checked against the source, with a link to it.
ThePrintAI productsAgents
Agoda's AI Assistant now lets people in 21 markets plan and book trips in 1 conversation, with live hotels, flights and tours shown on a trip planner canvas. It is a useful example of chat that ends in a real booking, where results move out of the thread into a shared, editable plan.
Related patterns:Generative UISide pane
9to5GoogleAI productsAI UX
Gemini now shows a Memory summary that groups what it knows about a person into categories, with a plain-language box to add or remove details and a history of changes. It is a clear, current model for memory controls: show what is remembered, let people fix it in their own words, and log every change.
Related patterns:Memory controls
Apiiro found 2,539 secrets in 1,994 public repos because approving a command with 'Yes, and don't ask again' saved it, password included, to a project file that then got committed. It shows how an approval shortcut can quietly store sensitive data, so remembered permissions need clear scope and a visible record.
Related patterns:Approval gateConsent and dataActivity log
9to5MacAI products
Apple now sells a one-time $49.99 Intelligence usage pack in Logic Pro, used only after the monthly limit runs out, with credits that expire after 12 months. It makes usage limits a visible part of the AI experience, so products need clear meters, warnings before the limit and an easy way to buy more.
Related patterns:Capability disclosure
AnthropicAI products
Anthropic released Claude Fable 5.2 for all users and Mythos 5.2 for vetted organizations, with replies up to 30% faster and an optional fast mode at a higher price. Speed and price now vary by mode, so model pickers need to explain each trade-off in plain words.
Related patterns:Mode and model
Tech My MoneyAgentsAI products
Perplexity's Teammates are always-on agents, each with its own desktop and browser, that post progress updates and ask for approval. People can inspect results, redirect a task or take over the desktop. A clear mix of approval gates, mid-run steering and takeover in 1 product.
Related patterns:Approval gateUpdate mid-runWatch and take over
9to5GoogleAI UXAI products
A Google Messages beta replaces the floating Gemini button with a small button in the top-right corner, next to the profile picture. Search also moves to a floating button at the bottom. It shows how a big app is rethinking where an AI entry point should sit.
Related patterns:Entry point
9to5GoogleAgentsDesign tools
Antigravity 2.29 adds Remote Agents that keep working in the cloud after the laptop closes, plus a Gemini Live voice mode. People can check on agents and grant permissions from a phone page. Automations are now sorted into Scheduled, Always on and Paused.
Related patterns:Long-running tasksScheduled agentsApproval gate
TechCrunchAgentsAI products
Meta's Muse agent now has Spaces, focused conversations in the sidebar that Muse builds into a tracker, planner or project board. Connected apps keep each space updated in the background. It moves the agent from a single chat thread toward a lasting workspace.
Related patterns:Agent workspaceGenerative UI
TechCrunchAgentsAI products
Invited US customers can text Alexa in plain language to find products, add them to the cart and buy, plus track prices and search by photo. A wider rollout comes in the coming weeks. No confirmation step is described, so the purchase moment is worth watching.
Related patterns:Approval gateMultimodal input
Jakob Nielsen's roundup covers Atlas, a research tool that shows a finished agent run as a tree of paths taken, set aside or ruled out. In a 14-person study it raised decision transparency and lowered workload versus a chat transcript. A strong case for richer activity logs.
Related patterns:Activity logExplanation
Nielsen Norman GroupAgentsResearch
Nielsen Norman Group published a definition of the agent harness: the software that runs the loop between an AI model and its tools. Its context, permissions and limits decide what users can ask an agent to do. A useful term for designers who need to talk about agent limits with engineers.
Related patterns:Capability disclosureAutonomy dial
Google Cloud blogDesign tools
Google made Gemini Omni 1.2 Flash generally available with clips up to 15 seconds, better reference following and a 30% price cut. Users can refine the same clip in conversation, swapping a product or changing lighting or camera angle instead of starting over. It is in Google Vids now and coming to Gemini for business.
Related patterns:Edit outputRegenerate and branchesAttachments
EngadgetAI products
Perplexity added Teammates to its Computer product: agents that each own a job, run 24/7 and notify you when a decision is needed. Each one gets its own sidebar entry showing what it is doing and what it finished, so you can redirect it mid-run. Max subscribers only, at $200 a month.
Related patterns:Scheduled agentsAgent workspaceUpdate mid-run
SkiftAI products
On its Q3 earnings call Airbnb said a specialised AI agent will reach guests in the first half of 2027, booking directly inside the app instead of linking out to partner sites. CEO Brian Chesky argued that specialised agents will beat general ones. A signal that big consumer apps are designing agents as native flows, not chat add-ons.
Related patterns:Entry pointIntent preview
The NeuronAI products
OpenAI's Images 2.6 lets ChatGPT turn a prompt or a still photo into an animated GIF. Users can open the result to change playback speed, edit single frames, or describe changes in chat, though ChatGPT cannot edit GIFs it did not make. Worth studying for how direct controls and chat edits sit side by side on one output.
Related patterns:Edit outputVariationsInline action
TechCrunchAI products
US users can now call or video chat with Meta's Muse agent on mobile, web and Mac. A default animated avatar reacts with expressions and gestures matched to the conversation, and it can be swapped for a person, animal or other character. It shows how much of a voice interface is really about turn-taking cues and a visible listener.
Related patterns:Voice inputVoice outputAI label
9to5MacAI products
Anthropic launched Personal Finance in Claude for paid US subscribers. Users link bank and card accounts through Plaid by signing in to their own bank, access is read-only, and Claude cannot move money. The flow is a useful reference for consent screens, clear scope limits and the reminders to check figures before acting on them.
Related patterns:Consent and dataDisclaimerCapability disclosure
Cursor changelogDesign tools
Cursor's cloud agent can scaffold a web app, add a database and sign-in, and host it. Nothing goes live until security checks pass and the user explicitly hits publish, and team admins can add review rules, approve blocked publishes or unpublish any app. A clear example of an approval gate placed at the one irreversible step.
Related patterns:Approval gateLong-running tasksErrors and refusals
Microsoft LearnAI productsAI UX
Microsoft 365 Copilot release notes say Excel chat replies now show changes in a new format, with links that jump to each one and new sheets, tables and ranges highlighted. Showing an AI's edits in place helps people check the work before they trust it.
Related patterns:Activity logExplanation
Nielsen Norman GroupUX industryAgents
Nielsen Norman Group argues that AI agents acting for people are widening UX beyond human users, and proposes "experience design" as the broader name. Designers now shape what both people and their agents meet. A useful frame for anyone designing pages and data that agents will read.
Nielsen Norman GroupUX industryAgents
Nielsen Norman Group warns that AI-native development can shrink UX to supplying context for coding agents. It argues UX must still find real user needs, shape what gets built, and test it before and after launch. A useful read for teams moving to agent-heavy workflows.
Microsoft Tech CommunityAI productsAgents
Microsoft's Copilot update brings Chat and Cowork under 1 Home with shared history, Projects that hold instructions and files, and Automations that run tasks on a schedule or event. Autopilot, an always-on agent with its own mailbox, reaches early customers with approval guardrails for consequential actions.
Related patterns:Scheduled agentsAgent workspaceApproval gate
iPhone in CanadaAI productsAI UX
Google added an in-chat editor to the Gemini app for all users. Files Gemini creates, from PDFs to Docs, Sheets, slides and code, open as a preview you can expand to type in, or you can highlight 1 part and ask Gemini to change just that. Mobile supports PDFs only for now.
Related patterns:Edit outputSide paneInline action
TechCrunchAI productsAI UX
Apple's iOS 27.2 lets Siri AI remember details from earlier conversations and adds French, Japanese, Korean, Portuguese and Spanish. Users can choose which apps Siri's memory leaves out. A big-platform example of memory shipped together with a clear control over what it may use.
Related patterns:Memory controlsVoice input
Nielsen Norman GroupResearchUX industry
Nielsen Norman Group surveyed UX professionals and found AI saves time on nearly every task. Yet mandates, unreliable output and inflated expectations eat much of the gain. Checking and fixing AI output is the top complaint, a reminder to design AI features that are easy to review.
Related patterns:Edit output
Anthropic opened a research preview of Claude Duet, a separate app where Claude sends a morning brief, checks connected apps every few hours and works on tasks in the background. It asks before it sends, pays, signs or deletes anything. A clear case of proactive nudges paired with an approval gate at the risky steps.
Related patterns:Proactive nudgesApproval gateScheduled agents
Android AuthorityAI productsAI UX
OpenAI is rolling out locked chats on every ChatGPT plan except Healthcare. Chosen chats hide from history, pins and search, open with a 4-digit PIN or Face ID, and keep a memory apart from regular chats. Resetting the PIN deletes them for good, a hard trade-off that needs very clear warning copy.
Related patterns:Memory controlsTemporary chatConsent and data
Nielsen Norman GroupResearchUX industry
Nielsen Norman Group compared AI-generated respondents with a survey of 604 people. Both pointed to the same conclusion, but the synthetic answers were far less varied. Fine for a quick direction check, risky when the edge cases are what you need to find.
TechCrunchDesign toolsAgents
At Config India, Figma launched Specs, a document on the canvas where teams write down their design language for Figma's own and outside AI agents to follow. Code layers also moved to open beta, so prototypes can be backed by real code. Both aim to make AI output more predictable inside a design system.
Related patterns:Structured prompt
Also as an RSS feed. Headlines link to the original publisher; summaries are written here in plain English.