Meta Muse 登陆新硬件，Amazon 表示不满

🔵 Meta’s Muse comes to your keyring - but Amazon isn’t happy Plus: Figma’s CEO says UIs won’t disappear, YouTube and Spotify ship algorithmic controls, OpenAI’s new mental health benchmark Rich Holmes Sep 27, 2026 4 1 Share Hi product people 👋, This week, Meta’s Muse agent crossed over 500,000 users and made the jump into an impressive looking new piece of hardware, but Amazon has banned it. We’ll take a closer look at the details. Plus, Figma’s CEO explains why he thinks UIs won’t disappear and the AI Agent safety concerns continue, prompting companies to ship new Agent auditing capabilities. This is your Product Briefing. Have a great week ahead! Follow on Substack Notes Key reads and resources for product teams Making Sense of the New Product Development Models Spotify doubled its AI assisted code in a year, but faster output created new problems further down the pipeline. I dig into this alongside similar signals from Linear, Ramp, PostHog and Square to map five core shifts reshaping product development. (Department of Product) How Dropbox built its new API documentation This week, Dropbox’s product team gave its API documentation a significant upgrade, and you can now test endpoints right in your browser without ever touching a separate tool. If you’re curious about what API docs look like in 2026 this case study is an interesting read. (Dropbox) Stripe’s CEO explains how Stripe uses Claude Code internally In this conversation, Patrick Collison explains how Stripe used Claude Code to build products like Stripe Projects in weeks instead of months and why “every codebase is now a prompt for another codebase”. (Anthropic) How Sentry built an internal evals platform from scratch and now runs 1,800 eval runs a month on pull requests Sentry started in May and now runs evals across 17 AI product surfaces using production-derived scenarios rather than synthetic tests, reporting 20-50% cost reductions and 10-20% eval-pass improvements. (Sentry) CAFE(S): Atlassian’s new framework for evaluating the context you hand to an agent Atlassian names five properties for context quality, clarity, actionability, fidelity, efficiency and security, and treats token spend on context as a real budget line. It also cites the EchoLeak prompt-injection exploit against Microsoft 365 Copilot as the first documented real-world exfiltration of this kind. (Atlassian) New product features and innovation this week Meta continued the aggressive rollout of its Muse Agent with the release of a bunch of new features and integrations. A new GitHub integration lets users review pull requests, manage issues and notifications, an Expedia integration lets Muse autonomously book hotels and travel logistics from a simple destination request and Meta’s AI chief Alexandr Wang confirmed that Muse is also getting its own email address. Alongside these new use cases, Mark Zuckerberg also announced a major new piece of hardware with the announcement of Muse Charm: Charm is a small tamagotchi-sized toy that can attach to your keyring to bring Muse to wherever you are. The form factor here seems pretty smart; previously, companies have attempted to ship devices designed to either replace your phone or force users to carry an additional clunky device alongside it. Most of these previous attempts have failed miserably (see Rabbit and Humane Pin), but in a world where most users still carry around their keys, shipping a small device like this that attaches to those keys feels like something users would be willing to do - and it could be the start of a new form factor we see from other companies like OpenAI who are working on a small device of their own. But, it’s not all good news for Meta. Amazon confirmed that the agent is now officially banned from making purchases on Amazon. Amazon says it was never told by Meta that it would access its stores and that Muse doesn’t identify itself correctly when it browses. Amazon lost a previous lawsuit against Perplexity’s Comet browser in August and so it’s unclear whether it will be successful in blocking Muse in the long term. Google launched Gemini 3.8 Live with Avatars, designed to bring a more conversational experience to users. Google says that by pairing near real-time video generation with speech, it enables product teams to expand their virtual offerings for use cases like customer service and interactive walkthroughs. Here’s some examples of it in action: New features you can use at work Atlassian shipped a new tool called Protobox for AI-assisted design and prototyping. This is a new Mac app, built over 18 months that embeds coded prototypes and agent skills to help product teams shape new ideas through prototypes. Atlassian says it’s already used internally across teams at Confluence and Loom. Atlassian’s VP of Core Design explained how they built it here. Linear built a dedicated bot that turns social media mentions into structured product feedback that can be used to shape new features or iterations of your product: UserTesting released new MCP servers for research recruitment across the UserTesting and User Interviews platforms. Product teams can recruit participants, launch studies and validate AI-generated concepts from inside ChatGPT, Claude, Figma Make or Gemini. Asana launched Command, a new planning system built for software delivery. It gives engineers and coding agents shared context, updates release forecasts dynamically as work changes and surfaces delivery risk without status meetings Perspectives from the world’s leading product leaders and operators Dylan Field, CEO, Figma, says that user interfaces will not disappear as agent adoption continues to grow, because humans need human-readable ways to audit what an agent did. He expects new input modalities like voice and VR to expand the design surface rather than shrink it, with design becoming “the new code”. Mike Cannon-Brookes, Co-founder and CEO, Atlassian, argues that software factories are mostly hype, comparing real progress to adding machinery to a chocolate factory station by station rather than a revolutionary leap. Humans stay essential for initiation, judgment and exception handling, and he warns against the “I do AI better than you” positioning that’s becoming more common. Cristina Cordova, COO, Linear, predicts that consumers won’t actually delegate spending or personal communication to agents yet, drawing a hard line between low-stakes tasks like restaurant bookings and access to credit cards and inboxes. She puts current agent payment capability at the novelty stage for anything higher-stakes. Tom Eggemeier, CEO, Zendesk, argues AI spend should be measured in cost per completed task rather than token price, describing an internal practice of routing different work types to different specialised models that cut average cost per task 22% in a month. Thariq Shihipar, Member of Technical Staff, Anthropic, says that product development teams should resist shipping 10x more features simply because models got more capable. Instead, he argues that teams should spend the time on user research and prototyping. 📈 Product data OpenAI has published a new mental health benchmark called MentalHealthBench and says that Astra GPT-6 scores the highest. OpenAI worked with more than 80 licensed psychologists and psychiatrists from 22 countries to build the benchmark and you can read the full methodology here but for product teams building AI Assistants that need a strong ability to navigate complex mental health issues, the paper is worth a read. Meta says Muse passed 500,000 total users, 250,000 daily active users and 2 million prompts submitted in its first week. Proton surveyed 4,014 chatbot users across France, Germany, the UK and the US and found 66% discuss sensitive topics such as finances and mental health with AI, while only 11 to 20% trust AI companies to protect that information and 80% want a chatbot built specifically for privacy. Claude Opus 5.5 is the best model for operational tasks, scoring 40% on Zapier’s AutomationBench which tracks 657 workflows. OpenAI’s new model, GPT-6 Sol, scored 33.2% on Zapier’s AutomationBench, up from 28.77% for GPT-5.6 Sol. For context, in April the highest scoring model was scoring 10%. Lovable’s annualised revenue reached $600 million, up from $500 million in June, with users at two-thirds of Fortune 500 companies and nearly a billion monthly views on apps built on the platform. LinkedIn reports that verified members get 50% more post engagement and nearly 90% more profile views, and is now piloting peer-to-peer confirmation of shared work or school history. Salesforce found agentic search, where a shopper’s first step is a question to an AI assistant, grew 200% year on year, with AI-referred traffic up 150-428% every quarter since 2024. Vercel’s Agent Skills registry hit one million skills and 280 million installs in just seven months, faster than GitHub reached one million repos (27 months), the App Store reached one million apps (63 months), or npm reached one million packages (117 months). Dropbox says that 70% of code at the company is now AI generated. Emerging new trends powered by Product Intelligence Here are some of the new emerging product trends uncovered this week, powered by Department of Product Intelligence. Paid subscribers can get access to the full datasets underpinning these trends with over 150+ data points here: Upgrade to unlock Intelligence Reports 1/ Users are being given more control over their algorithm and UX YouTube shipped a major new feature this week which gives users the ability to craft their own custom feeds, reflecting a growing trend of users being given more control over the algorithms that power the products they use: Spotify has also officially expanded its Taste Profile to users in the US which lets users directly shape their recommendation and homepage feeds and Meta’s AI Chief said that Muse’s new Feed could be a sign of things to come where personal agents shape feeds based on their rich memory of conversations, comparing it to a “good friend” recommending content rather than traditional algorithmic systems. 2/ Token spend efficiency is top of the agenda - with companies flexing their ability to reduce costs Vendors are increasingly competing on token efficiency as companies look to reduce their AI token spend. AWS has released an open-source agent harness which it claims reduces token costs by 28% and Stripe has become the latest company to adopt the WebMCP standard - which it says reduces agent token cost by 42% during payment flows. You can read more about WebMCP and why it matters to product teams here. Elsewhere, Cursor says it can now offer users a 7% reduction in token costs thanks to tighter prompts, better caching and compressed reads and Linear added adaptive routing to its Coding Agent product this week, where it automatically delegates simpler tasks to faster models, keeping more expensive reasoning models for more complex work alongside open source. Leaders are also starting to share their own perspectives in managing token spend. Harvey’s co-founder said the company went from negative 50% to positive in a single quarter through model routing, harness improvements and training and Databricks’ VP of engineering said that giving 3,500 engineers access to GPT-6 Astra actually increased token spend by 60% - but only on complex work. They now run cohort-based experiments on smaller groups of engineers who are given a budget to experiment with model choices: 3/ Keeping AI agents under control is emerging as a distinct problem - and product category of its own AI agents are continuing to run wild with more attacks revealed this week. OpenAI agents were found to have breached the US Commerce Department, the SEC, and attempted access to the Education Department, plus Australian government health databases and university systems dating back to at least March 2026, with one successful compromise of Australia’s national healthcare system in June that OpenAI did not learn of until August. Meta’s Muse Mac app had two separate security incidents, including one in which allowed a local code execution to hijack the agent via an undocumented setting. As a result of these ongoing concerns around AI agents, platforms are scrambling to release new features designed to help product teams observe the behavior of AI agents. This week saw the release of Agent Observability in Snowflake, a new Workflow history feature in Attio and a new tool from AWS to help teams observe AI agents in applications. The Briefing is your curated report on what happened in tech and AI this week - hand crafted for product teams. Paid subscribers get the full DoP Substack including: The Knowledge Series for hands-on AI tutorials plus DoP Deep dive and Intelligence reports for in-depth analysis to learn product lessons from the world’s top tech companies. 4 1 Share Previous