Claude Duet arrives, Walmart opens its checkout to agents, and the $0.50 dashboard

Walmart Agent Gateway — lets AI agents buy for customers on Walmart and Sam's Club from next month, through a beta gateway built on the Model Context Protocol that keeps the shopper's history and personal data away from the agent. Purchases need approval by default, or users can allow spending within a set budget and review every agent purchase separately in an Agent Profile, switching the agent off anytime. Agent commerce just became a real channel: Amazon blocked Perplexity's and Meta's shopping agents, while Walmart is publishing terms agents must follow instead — no acting beyond the shopper's intent, no changing payment methods, and rule-breakers get blocked. If you sell anything, this is the template for your agent-facing surface. Users can pre-approve shopping within a budget, such as up to $100, instead of approving each purchase. theverge.com

Dessa — is a speech-to-speech model from a small lab inside Block that says it was trained from scratch on human conversation only — no text, no pretrained models — so it hears and speaks emotion, sarcasm, laughter and timing, and listens while talking like a person does. This is a research demo, not a shipped feature — waitlist only, no paper, weights or API yet — but it points at on-device voice agents that run locally instead of shipping your audio to a cloud. If it holds up, voice UX expectations move: interruption, backchannel and emotion become the baseline, not a nicety. Block says it runs faster than real time on a MacBook Pro M4 Max or better, was trained on a single NVIDIA GB200 NVL72, and is giving safety researchers early access. block.xyz

Record for agent — turns a short screen recording into a prompt an AI agent can act on: Loom captures your narrated transcript, screenshots of the elements you clicked, and the links you visited, then hands the package to any agent — or files it as a Jira work item assigned to one. For PMs this is the fastest spec workflow yet: record the current behaviour and the edge cases, point at the exact UI elements, and a coding agent receives a richer brief than most written specs produce. The recording then doubles as review evidence, so the team can judge the result against what was asked. Open beta on Loom Business + AI and Enterprise plans, via the new Loom for Mac or the Chrome extension; a 7,000-person survey found 70% said a short recorded video would have prevented their last miscommunication. atlassian.com

1Password for Agents — lets an AI agent sign in to your accounts through a broker: the agent proves its identity with a digital key, you approve on your own device, and 1Password fills the login so the credential never enters the agent's context. It works with Meta Muse, ChatGPT Work, Dots, Claude Duet, Grok Bot and Town at launch. This is becoming the default trust pattern for any agent feature that touches user accounts — no pasted passwords, per-task approval, no standing access. Expect users to ask why your agent needs their password when 1Password's brokers don't hand one over. Limits worth quoting in your risk review: a signed-in hijacked agent can still act inside the account, and the approval text comes from the requesting agent. Pricing wasn't stated in the launch coverage. explainx.ai

Trello MCP — connects Trello to assistants like Claude, ChatGPT, Gemini and Cursor, so anyone can create boards, lists, cards and checklists by asking — turning a planning conversation directly into trackable work instead of twenty minutes of manual re-entry. The signal is who it's for: Atlassian says nearly half the people using its MCP tooling are not on software teams. Agent features are now a mainstream-workflow expectation rather than a developer feature, and whatever planning or productivity tool you ship is now an integration target for someone else's assistant. Available to all Trello users via trello.com/mcp; Trello counts more than 100 million signups. atlassian.com

Pragmatic Engineer AI-spend survey — surveyed 27 workplaces on AI spend and pull-request throughput, plus aggregate data from CodeRabbit and Gitar. Company size and spend turned out to be unrelated — a two-person .NET shop spends $3,000 per developer a month while a 1,500-engineer public firm caps at $150 — and weekly PRs among developers using AI code review roughly doubled year over year. Two planning facts: AI spend is now a real budget line with caps and justification forms (one 2,000-dev quant shop hit $8M in September, 16x its January bill), and one staff engineer reports PM-facing prototypes that used to take a week or two now land in a few hours. Scope your discovery and prototyping plans to that speed, and put a per-seat AI line in next year's budget. Spend peaked in summer and has fallen since, roughly halving for companies on API plans as they moved to Opus 5.5 and open models. newsletter.pragmaticengineer.com

Three engineers quietly replaced a third of their Looker. At a 500-person company paying about $150k a year for Looker, three data platform engineers built a Claude-generated HTML dashboard tool as a side project. Within eight months it had as many weekly users as Looker, Looker's usage fell by about a third, and the whole thing cost 325 commits and roughly $6.4k of Claude tokens — against about $100k of licence fees over the same months. The build-versus-buy math for anything that is essentially UI over data has collapsed: a dashboard now costs about $0.50 in tokens. That cuts both ways for your roadmap — internal tools you rent are suddenly questionable, and product features that are dashboards under another name are suddenly cheap to prototype and vulnerable to customers vibe-coding their own. Each dashboard takes the analyst agent about 1 million tokens; the cost fell from about $2.50 to about $0.50 after a move to Opus 5.5. david.lequin.dev

Atlassian developer survey — of 3,500 developers and managers across six countries found 99% now report time savings from AI tools and 68% save more than 10 hours a week — but half still lose 10 or more hours a week to non-coding work like finding information, adapting to new tech and context switching. The number to plan around is the empathy gap: the share of developers saying leaders don't understand their pain points jumped from 44% to 63%. Banking the AI hours saved without removing the friction points gets you the unfair-expectations trap the report warns about, not faster delivery. Developers spend only 16% of their time actually coding, and coding was never a reported time-waster. atlassian.com

Spotify — said on its Q3 earnings call — 788 million monthly users, nearly €5 billion revenue — that its AI bet is to be the media service any agent talks to: first to integrate with Meta's Muse, plus ChatGPT and Claude. The agent asks Spotify, and Spotify reasons over the request without sending listener data to the partner. Against Amazon's blocking stance this is the counter-playbook for platform products: expose an agent surface that leans on context only you have — listening history going back to 2008 — rather than fearing disintermediation by the assistant layer. Co-CEO Gustav Söderström frames agent growth as a tailwind: people will ask agents to build playlists, not to summarise songs. techcrunch.com

Meta Muse — confirmed more than 10 million users on Meta's earnings call, with about 25% of people who set it up using it daily. Over a quarter of its engaged shopping users averaged purchases above $200 — activity Zuckerberg said even Plaid hadn't seen from any other connector. These are the numbers to reach for when someone asks whether anyone actually uses consumer agents. And note the monetization model: keep it free for most, take a percentage of commerce, charge power users for advanced models — the same shape agentic commerce pricing will likely take everywhere. Meta expects to improve Muse's cost efficiency by more than 10x within a year, after which it expects the app to be profitable at revenue scale. techcrunch.com

Agent Skills — is being donated to the Agentic AI Foundation — name, logo, GitHub organisation and agentskills.io domain — the same neutral home as the Model Context Protocol, with maintainers from Anthropic, Google and OpenAI acting as individuals. Anthropic counts more than 70 agent platforms using the format, 1 million published skills and nearly 280 million installs. If you were waiting to see whether skills were a Claude feature or shared infrastructure, that question is now answered — building on the format is a governance decision, not a vendor bet. At AGNTCon the foundation also took in OpenAI's Open Responses, Hugging Face's OpenEnv and Alibaba's OpenSandbox. The adoption figures are Anthropic's own with no stated counting method, and a Snyk audit earlier this year found 76 malicious skills among those scanned — security guidance is on the new roadmap, not yet in the spec. simonwillison.net

Google — A privacy-policy change effective October 16th adds an exception: if you connect Drive, Gmail or Photos to Search's Personal Intelligence AI features, that content can be used to personalise ads, which the old policy promised it never would. Matters because personalization features now come with an ad-revenue trade-off your users will ask about — factor it into the consent copy for anything that reads user content. Users can disconnect apps from Personal Intelligence or disable personalised recommendations in Search settings. theverge.com

AI Deceptive Advertising Act — Governor Shapiro signed the law on Thursday: any ad significantly modified or generated by AI must carry a label in the ad's own format, regardless of where the ad was created. A compliance line for any team shipping AI-made marketing assets nationally — ads aired in Pennsylvania carry the rule with them. It carves out parody, satire and evidently fictional content, and "significantly modified" is left subjective. theverge.com

Legora — The Swedish legal AI company raised a $550 million Series E led by Atreides, with ARR above $200 million — up from $100 million in April — and 140,000 legal professionals using it across 80 markets. Matters because vertical agents are pulling real enterprise spend fast: rival Harvey raised the same amount at $15.5 billion a month earlier, and the two shipped near-identical Intake products a day apart. Customer Strabag is doubling licenses from 200 to 400 this year, reporting up to 80% time saved on data processing agreements. trendingtopics.eu

Orchestra Bio — OpenAI acquired the four-year-old startup, whose agents track drug-development operations like trial scheduling and manufacturing batches, and is folding the team into its Rosalind life-sciences line; terms were not disclosed. Matters because the labs keep buying the workflow layer above the model — the pattern of agents monitoring processes and handing decisions to humans is spreading to every regulated industry, including yours. Rosalind's named partners include Moderna and Novo Nordisk; OpenAI published no customer results or product timeline for the combination. techcrunch.com