AI Brief, November 20: Apple's agent hire, Gemini on Mac, Muse privacy leaks
Apple hires for a computer-using AI agent, Gemini gets computer use on macOS and Tool Retrieval, and Meta's Muse is reported mixing up user data.
By Sean WeldonAgents moved closer to the operating system today, with Apple reportedly hiring to build an agent that can drive a computer the way a person does. Google shipped computer use for Gemini on macOS and a Tool Retrieval preview for large tool libraries, while Reuters reporting on Meta's Muse showed what happens when agent context isolation fails.
Apple job listing points to an in-house AI agent that can use a computer
The Verge reports that Apple is working on its own AI agent that can browse the web, work across apps, write code, and call services, based on a job listing spotted by Bloomberg for an engineer to lead an agentic AI assistant. The listing says the agents "plan and carry out multi-step tasks across apps, browsers, command lines, code, and APIs" and can coordinate specialized sub-agents, with the role working across Apple's Siri, UI Frameworks, and Apple Foundation Models teams. TechCrunch reports Bloomberg also noted the work remains in early stages, while macOS 27.3 will warn users against granting third-party agents full access.
Why it matters: If Apple ships a first-party agent while tightening full disk access for third parties, builders targeting the Mac may need to plan around Apple's own agent APIs rather than their own computer-use stack.
Sources: The Verge AI / TechCrunch AI
Gemini API adds Tool Retrieval to keep large tool libraries out of context
Google launched Tool Retrieval in preview in the Gemini API's Interactions API. Developers add a tool_search tool and set defer_loading: true on tools they want held back: the model still sees each deferred tool's name and description, but parameter schemas load only when needed, inside the same request. Google says there is no index to build or host, search runs server side, and deferred schemas stay out of the prompt prefix so prefix caching still works. It also supports remote MCP servers and a client-side mode with "execution": "client" for your own registry or permission filters. Tool Retrieval works with Gemini 3.6 and later, including all Gemini 4 models, and Google says there is no separate charge.
Why it matters: Agents wired to dozens or hundreds of functions can cut input tokens and improve tool selection without building a retrieval layer, though Google advises measuring the savings against your own traffic.
Sources: Google AI
Gemini's Mac app gains computer use for local files and desktop apps
TechCrunch reports Google brought computer use to the Gemini desktop app on macOS, letting the agent navigate native Mac apps and browse websites in the background to complete multistep tasks across local files, apps, and the web. Work appears as a live preview in the chat while the user's screen and cursor stay free, and users can pause, stop, and approve actions and choose which apps Gemini can access. Google suggests uses such as filling out forms, repeating a task across a batch of files, organizing files into folders, and creating slideshows. It is available in English in the U.S. and requires the Gemini Mac app plus a Google AI Plus, Pro, or Ultra subscription.
Why it matters: OS-level agent automation is now a consumer feature from three major labs, raising the baseline for what desktop AI products are expected to do.
Sources: TechCrunch AI / The Verge AI
PostgreSQL 19 ships read-your-writes for standby reads via WAIT FOR LSN
Simon Willison notes that PostgreSQL 19 arrived two months later than the usual September cadence, which he attributes mainly to issues the team had to address during what he calls the ongoing AI-reported vulnerability apocalypse. His favorite feature is read-your-writes consistency for applications that send reads to standbys: a new WAIT FOR LSN command makes a standby session wait until it has replayed the write-ahead log to a given log sequence number before running a query, with an optional timeout that errors if the replica does not catch up. He shows capturing pg_current_wal_insert_lsn() after a write and passing it back on a later request, and compares the pattern to MediaWiki's hand-rolled version on MySQL replication.
Why it matters: Teams scaling AI backends with read replicas can get consistent reads after writes from Postgres itself instead of maintaining replication-lag workarounds in application code.
Sources: Simon Willison
Meta's Muse agent reported mixing up conversations and exposing private data
TechCrunch reports, citing Reuters reporting on internal company posts, that Meta employees raised concerns that the Muse agent has been mixing up conversations across accounts and exposing private user data. Employees said Muse surfaced childcare requests they never made, career advice they never asked for, product orders they never placed, and bank balances that were not theirs, including an unprompted message listing two bank balances during a chat about weather and tides. Another employee said Muse reminded her to reply to a Signal message from her sister and called it an extreme privacy violation. Meta told TechCrunch employee testing is core to product development and that employees use different versions of Muse than the public.
Why it matters: Context and session isolation is the failure mode to test first in always-on agent products, because cross-account leakage surfaces as plausible-looking content rather than an obvious error.
Sources: TechCrunch AI
Architect Labs says AI designed and verified Frostbite, a new CPU architecture
Architect Labs introduced Frostbite, a general-purpose CPU with a new instruction set that the company says its AI systems discovered, designed, and verified end to end, iterating autonomously over the compiler, ISA, architecture, microarchitecture, RTL, verification, and final GDS. The design exploits interaction nets, where every step is local, every dependency explicit, and every value has one owner. On 25 workloads, 15 of them held out until final evaluation, the company reports a 4.59x performance-per-watt multiple against a modern silicon-proven dual-issue in-order superscalar core. Agent swarms worked against a machine-checkable specification language, ASL, that translates to Lean; across four searches and a final round of ablations the swarm tried 58 ideas and kept 37.
Why it matters: The notable part for builders is the method: a formal, machine-checkable spec as the single source of truth, which the company says prevented a class of agent hallucinations and reward hacking in large swarms.
Sources: architectlabs.com / Hacker News discussion
Claire Vo winds down ChatPRD free trials after token fraud, Simon Willison reports
Simon Willison writes up a video from Claire Vo, who runs the AI product platform ChatPRD, describing a flood of token fraud from groups abusing the product's free trial. Some abusers wanted free AI usage, while others resold tokens or ran commercial operations on top of the ChatPRD agent, including one Vo describes as a Roblox and Fortnite map studio looping free trials and promo codes to build maps through her agent. Willison notes a common prompt running through the agent involved SVGs of pelicans riding bicycles, which he suggests was used to fingerprint the underlying model. Vo is winding down free trials entirely.
Why it matters: Anyone offering a free tier on a metered AI product is effectively underwriting resale, so abuse detection and trial limits belong in the launch plan rather than after the bills arrive.
Sources: Simon Willison
How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.
Sean Weldon builds apps and custom software at 22Atlas.