United walls off AI agents; Google prices decisions at a dime; Grok share links leak

🔥The one thing to know

🧱New ground broken

• Gemini Decisions API — a new Google endpoint, opened today, that answers closed-set questions — yes/no, multiple choice with up to 255 options, or rubric scoring — over text, images, audio, video and PDFs in a single pass, returning probabilities and confidence scores instead of prose. This is the routing-and-policy plumbing teams used to hand-roll around a chat model. Bounded decisions are the cheapest place to put an agent, and Google just made them a line item: a triage, moderation or routing feature can now ship as a list of options rather than a custom classifier. Google already uses it to route queries inside Gemini Notebook. Answers land in 150–250 ms end to end; one request can carry up to 1,000 questions, input costs $0.10 per million tokens, generated output costs nothing, and AI Studio offers a free tier for prototyping. Google blog

• Apple Harbor — a leaked, unannounced Apple app that would turn a Mac into a private server for open-weight models — downloading and running them locally, and pooling the memory of two or more Macs over Thunderbolt 5 so models too big for one machine can run. If it ships, private on-device AI stops being a hobbyist setup and becomes an enterprise checkbox: regulated workloads served from Mac hardware with zero cloud egress. It is a leak via Aaron Perris — Apple has not announced it, and it is reportedly in private preview under its Claris subsidiary. The hardware it targets is not cheap: Engadget puts a four-Mac-Studio setup running Meta's 816-billion-parameter Muse Spark Open at about $38,000. 9to5Mac

🚀Agent features and products that impact product management

• Databricks Genie — an enterprise agent that takes a goal and works unsupervised across a company's own data. CEO Ali Ghodsi demoed his real instance dissecting the worst-scored parts of an employee survey, rating executives' proposed fixes, and summarizing a failed sales pitch. Read the fine print before quoting it: Genie and Genie Ontology are generally available, but the always-on mode, Genie Search and Ontology Manager are only 'coming soon' to preview. Pricing is usage-based, with spend caps set through Unity Gateway. Databricks claims Genie answered correctly in 91% of task runs in a benchmark built with Turing, on under half the tokens of a stronger frontier agent. TechCrunch

• Google Tabletop — a Google Labs experiment where a group describes a tool — trip itinerary, dinner poll, expense splitter — and @tabletop builds it as a shared space everyone uses and edits together like a Google Doc; members can ask it to change the rules later without starting over. It's the consumer face of prompt-generated mini-apps with multiplayer editing, and Google itself compares the look to Asana and Monday. Labs experiments like NotebookLM have a record of graduating into full products. It's a waitlist beta — US-only, 18 and over — and Google hasn't said which model powers it. The Verge

🧭How product teams work now

• AWS Emulator — a free AWS tool that simulates more than 200 of its cloud services on a local machine. The notable part is who wrote it: AWS says AI agents produced all the code, no engineer wrote or reviewed a line, and the company presents that absence as the selling point. The planning signal: a major vendor treating zero human code review as a marketing point. Expect 'who reviewed this?' to become a standing question in your own release process. GeekWire via Techmeme

• OpenAI Codex — at the AIE Code Summit, OpenAI's developer-experience engineers described how they now work: publish one cloud environment and run many isolated tasks in it, steer agents by voice, and start jobs from their personal dot because it already holds the context. The shift worth copying: work is managed as threads you check in on, not prompts you babysit. Kundel just asks local Codex to hand a task to the cloud; Guo uses voice to manage his threads, not to dictate code. Guo's three uses for dots: safety net, delegation sink (used heavily in Slack), and presence everywhere you work. Latent Space

• HN stress-tests the 100% AI mandate, a year in. a Hacker News thread on a post claiming that a year into a 100% AI coding mandate, nobody at the company understands its codebase. Commenters doubt the post itself is human-written, but the substantive argument is about who pays when the company gains velocity and the engineer loses learning. The recurring report in the thread: an engineer at a large F100 firm spends ten-hour days prompting Claude and learns nothing. That's a review-capacity and retention risk that shows up in your estimates, not in velocity metrics. reddit.com · from the Hacker News discussion

📰Market and platform news

• Grok Build — security firm Claroc found Grok Build puts the random words of its share links in the subdomain, so passive DNS logs make them enumerable: of 43,601 grok.me links, nearly 95% opened without sign-in, including 24 apps built by SpaceXSI's own staff. If your team ships agent-built internal tools behind share links, this is the failure mode to design against: the default was 'anyone with the link,' and DNS made the links listable. Claroc estimates 650–800 apps held private data, including an ICU patient briefing and plaintext passwords; the report shows exposure, not proven misuse. Ars Technica

• OpenAI's quarterly revenue reportedly passed $10 billion. Bloomberg, citing unnamed sources, reports OpenAI's quarterly revenue passed $10 billion for the first time, up almost 50% from $6.7 billion in Q2, with its adjusted operating margin improving from -150% to -109%. Unconfirmed, and not a company filing. The number a CFO will quote back: the market is expanding fast enough for both leaders to grow at once — Anthropic says its own annual run rate reached $81 billion by end of October. OpenAI's annual run rate is reported at $50 billion; the two companies count partner sales differently. TechCrunch

• Pomerado — a YC F26 launch that turns any website into a deterministic, self-healing API an agent can call: describe a workflow, its agent performs the task once in a browser and generalizes it into an MCP, repairing the tool when a site changes. This is the middleware that fills the gap when partners won't build agent APIs — and, given United's move above, the next thing partners will try to block. The founders are ex-McKinsey and Scale AI. The team reports a 2.2% error rate with 82% auto-repaired; it's MIT-licensed with a free hosted version, and dealership users switched from computer-use flows that took over an hour. Hacker News

• Claude Fable 5.2 in GitHub Copilot — Anthropic's newest frontier model is now generally available across GitHub Copilot tiers — Pro+, Max, Business and Enterprise — where early testing had it solving more tasks than Fable 5.1 with fewer steps. Business and Enterprise admins must enable the model policy; it is off by default. The catch that reaches a roadmap: Fable 5.2 requires data retention by default to run Anthropic's safety classifiers, so the zero-retention posture most enterprises assume does not hold for this model. GitHub Blog

🌍The big AI stories today

• Nvidia — Nvidia's board is reportedly preparing to oust Jensen Huang after a second reported affair — shares fell about 8%, so chip-supply plans just got less predictable. TechCrunch

• AlphaGenome 2 — DeepMind shipped AlphaGenome 2, a DNA model that nearly doubles causal-variant recall per cell type — relevant only if biotech sits on your roadmap. Google DeepMind

💡One takeaway