Claude Code refuses game work after a silent rule change; Jira makes agents measurable; OpenAI's round shrinks to $20bn

Waveshare ESP32-S3 1.75C — is a $50 devkit, officially sanctioned to talk to Meta's Muse AI cloud, that a Verge editor with zero coding experience rebuilt into a personal AI gadget over a few Claude Code sessions: conversation history, spoken replies via an ElevenLabs voice, a remote shutter for his phone camera, and a Tamagotchi mini-game with four levels of hunger and grime. Hardware customization by non-coders stopped being a firmware-team project. If your product has any hardware or devkit angle, the expectation that users reshape the device themselves is now the baseline — and the bar for what a shipped consumer gadget must beat. This is a DIY build by one journalist, not a shipped product; Meta's actual Muse Charm hasn't shipped yet, and he kept the device on guest Wi-Fi after a reviewer's home address leak. The Verge

Jira Autumn release — gives agents their own identity on boards, lists and work items through AMP, Atlassian's human-agent collaboration protocol, and adds an agent usage dashboard in the space Reports tab that ties agent activity — Rovo, coding tools, third-party agents — to cycle time, throughput and shipped work. It's on by default with no configuration, rolling out now and generally available across Jira Cloud by November 2026; custom templates also drop down to Standard and Premium plans. If leadership asks what agents actually do to delivery speed, this turns that question into a shipped report instead of a data project — and it sets the bar every competing work-tracking tool will be measured against this quarter. The dashboard needs zero setup and is switched on for every space by default. Inside Atlassian

Atlassian's agent-sprawl playbook arrives with customer numbers. Atlassian's argument is that agent sprawl isn't shadow IT but unapproved actors using your data and permissions to act on your behalf — and the product answer is permission inheritance, per-agent scoping, and Agent Insights dashboards so admins can see how many agents exist, who owns them and what they touch. The proof points are DocuSign's: ticket drafting fell from about 20 minutes to 5, backlog grooming from 2 hours to 30. This is the governance conversation your security review will have this quarter, with vendor numbers attached — useful for scoping what agents may do before sprawl does the scoping for you. Gartner projects the average Fortune 500 will run over 150,000 agents by 2028; only 13% of organizations think their governance is ready. Inside Atlassian

Arca — is an AI-native wealth manager where agents carry long-horizon analysis that can stretch over ten days of stop-start work, drawing on an internal memory system that tracks state, next actions and owners for every client. The division of labor is the instructive part: agents analyze and propose, a human advisor executes trades, and the hard limits are deterministic code the model can't argue with. This is a shippable pattern for regulated products — scope the agent to analysis, hard-code the no-go actions, keep a human as the actor of record — rather than waiting for model-level trust that won't arrive this quarter. The firm just closed a $48.5M round with General Catalyst leading and Index Ventures and Venrock participating. Latent Space

“agents prompt people more than people prompt agents”

Community Wisdom 304 — is this week's Lenny's Newsletter digest of its members' Slack, and one thread lands squarely in your beat: product people trading notes on what a harness for AI-driven development should contain — the scaffolding of specs, context and feedback that lets coding agents do real work instead of guessing. The same edition covers what most interrupts PMs and whether case studies beat behavioral interviews when hiring junior PMs. The harness discussion is the practice-side twin of what Arca and Atlassian are shipping: PMs are now expected to design the environment agents work in, not just the product. The full threads are behind the paid subscription. Published October 31 as a paid-subscriber-only edition. Lenny's Newsletter

Anthropic says a tampered Claude Code extension was published through Cursor's marketplace. The follow-up to yesterday's supply-chain story: the company says the booby-trapped extension reached developers through Cursor's marketplace rather than anything it published, and that nothing on its side — infrastructure or staff machines — was breached. It is handling the aftermath together with Cursor. For any team installing dev tooling via third-party marketplaces, this settles the threat model: the marketplace, not the vendor, was the weak link. Anthropic's official distribution channels were untouched, per its statement relayed by Felix Rieseberg. Techmeme

Meta has a humanoid prototype it doesn't plan to sell. Meta put together a working humanoid with Chinese robotics firm Fourier in about three months and is already on a second version, according to a sourced report; the company says it has no intention of selling these units. Cheap, fast robotics prototyping inside a big platform is worth tracking even without a product behind it — if Meta ever points a roadmap at it, the category compresses quickly. The report is by Jyoti Mann for The Information, relayed via Techmeme. Techmeme / The Information

Watermarking may cost output variety, not just polish. As major labs prepare to watermark AI-generated text to meet the EU AI Act's transparency guidelines, an analysis argues the no-quality-loss claims only hold for a single response: across several answers to the same prompt, watermarked outputs converge on similar phrasing. In one reproduction of Google's SynthID-Text method, four of five watermarked answers opened near-identically. If your product regenerates copy, offers retries, or samples several options and picks the best, plan for visibly less variety once watermarking lands — the diversity loss shows up exactly in those flows. Anthropic says watermarking doesn't hurt Claude's quality; its own cited paper describes a diversity-versus-detectability trade-off. Keyvan's blog

OpenAI funding round — OpenAI is now aiming for $20bn after securing about $11bn against a $30bn ask, at a $1.25tn valuation — a cooling funding climate that vendor-dependent roadmaps should quietly price in. Trending Topics

ChatGPT share-link flaw — A public ChatGPT share could run attacker code in the opener's signed-in session and read private chats; fixed October 26 and rated P1 — share links are an attack surface your product probably has too. Parameter

John Ternus on AI — Apple's CEO says the company won't ship AI for its own sake and doubts agents make apps unnecessary — the largest consumer platform remains deliberately unpersuaded by agent-first roadmaps. TechCrunch