Thu, Oct 22, 2026Thursday, October 22, 2026 · 20 stories · 8 min read

Claude Duet joins the always-on agent race with approvals built in

Ramit KoulFounder, Software Engineer & Innovator · Published 5:02 AM ET

Anthropic has opened a waitlist for Claude Duet, an agent that keeps watch over a user's apps and asks before it sends, pays, signs or deletes anything.^{1, 2} At GTC Berlin, Nvidia made a similar argument at the infrastructure level: its OpenShell Reference Architecture enforces agent limits from outside the agent, and Red Hat and ServiceNow are already building on it.^{3, 4} Distribution is also widening, with OpenAI models coming to Google Cloud and Walmart giving outside shopping agents their own profiles and spending caps.^{5, 6} Here are the 5 stories that matter most, then 15 briefs. Numbers in the text link to the references at the end.

Products

Anthropic opens a waitlist for Claude Duet, a proactive agent built on Opus 5.5

Anthropic has launched Claude Duet as a research preview. It is a separate app where Claude works as a proactive agent: it raises what needs attention, takes on tasks and keeps working when the user steps away.^{1, 2} Duet runs on Claude Opus 5.5 and is aimed at professionals and small business owners. Each morning it sends a brief on the connected apps, then checks them every few hours.^{7, 8} It has its own cloud virtual machine with a browser, so tasks keep running when the user's laptop or phone is off.²

Duet uses the memory, skills and connectors from the user's existing Claude account. It can also search past chats, and it supports plugins as well as Stripe, Link and Shop Pay.^{1, 9} Every purchase needs approval and uses a one-time card that works for that purchase only.⁹ Duet never sees passwords or sign-in tokens directly. A dedicated classifier and safety system prompt are meant to stop it leaking sensitive data such as card numbers.^{10, 8}

Access is limited to Pro and Max subscribers in the US. There is no extra charge, but background checks count toward plan usage limits.^{1, 11} During the preview, users cannot pick a model or watch the agent's browser. Anthropic also says Duet may still misjudge what is urgent.^{9, 8}

Why it matters for builders

Duet treats approval gates, per-purchase cards and periodic checks rather than constant polling as defaults, which is a useful reference for anyone designing agents that act for users.^{10, 9} Builders whose products depend on Claude connectors should expect more of that traffic to come from background agents rather than chat sessions.¹

Funding & deals

OpenAI models are coming to Google Cloud's Model Garden

OpenAI and Google Cloud are partnering to bring OpenAI's frontier models and Codex workflows into Google Cloud's Model Garden, sold through Google Cloud Marketplace.¹² Enterprises with Google Cloud spend commitments can draw them down for these purchases. Google lists GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna as the models it is adding.¹³

General availability is planned in the coming weeks.¹² With this deal, OpenAI models are on all three major US clouds. Amazon began hosting them in April, after Microsoft's exclusive license ended.⁵ Google says it is now the only hyperscaler offering Gemini, Claude and GPT on one platform.¹³

One analyst said Google's advantage will depend on data, governance and cost handling once every cloud carries the top models.¹²

Why it matters for builders: Teams already on Google Cloud will be able to test GPT models next to Gemini and Claude without signing a separate contract or moving data, and pay for them out of existing commitments.¹²

Industry

Nvidia's DSX POD packages a Vera Rubin AI factory into containers

At GTC Berlin, Nvidia introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers and validate before delivery.¹⁴ Nvidia says deployment can shrink from six months to 14 days once site utilities are ready.¹⁵ Each pod holds seven Vera Rubin NVL72 racks and arrives cabled, plumbed and tested, with delivery set for early 2027.¹⁴

Jensen Huang framed the product around lost revenue. By his estimate, a 1-gigawatt factory delayed by a quarter forgoes about $10 billion.¹⁴ Partners for the factory-built design include Bull, Dell, HPE, Lenovo, Schneider Electric and Vertiv.

Nvidia also expanded its DSX platform. Warm-water cooling is estimated to save 54 million liters of water a day at 2 gigawatts, and DSX Flex automates how a data center throttles its load when the grid needs it to.¹⁴ The 14-day figure assumes a ready site, and containers do nothing to shorten grid interconnect queues.¹⁴

Why it matters for builders: Neoclouds and sovereign AI projects in the hundreds-of-megawatts range could get capacity online sooner, which affects when new GPU supply reaches the market in Europe.¹⁴

Models

Recursive releases Flint, an open-weight local coding model built largely by its own AI research system

Recursive has released Flint, a mixture-of-experts coding model with 40 billion total parameters and 2.5 billion active per token. The weights are on Hugging Face under Apache 2.0.^{16, 17} The company says a small team of people and agents built it from scratch in under four months, with its Eureka Machine helping across data, training recipes, infrastructure and inference.¹⁷

Recursive says Flint leads MoE models under 50 billion parameters on SWE-Bench Pro v2, Terminal-Bench 2.1 and SWE-Bench Verified.¹⁷ In its own tests, Flint generated about 131 tokens per second on an M4 Pro MacBook Pro, roughly twice the speed of similar models, and passed 150 tokens per second on Nvidia's RTX Spark.¹⁶ The company credits Eureka with a sparser architecture, low-bit training that shrank the model by 71% and custom kernels.¹⁶

Wired's tester ran Flint on a 48 GB Mac and found it fast and accurate on small programming jobs.¹⁸ Recursive says Flint needs more guidance than large paid assistants and suggests splitting big tasks into smaller steps.¹⁶ The startup has raised $650 million at a $4.65 billion valuation and wants Eureka to reach full recursive self-improvement under human oversight.¹⁸

Why it matters for builders: Flint gives teams a free model for private, offline coding agents on a single workstation, with no usage fees.¹⁶ It is also an early public test of whether automated research pipelines can produce competitive models with very small teams.¹⁸

Infrastructure

Nvidia's OpenShell Reference Architecture moves agent enforcement into hardware

Nvidia has launched the OpenShell Reference Architecture, a design for building, containing, observing and enforcing safe agents across enterprise infrastructure and edge systems. It is available now.^{3, 19} Models run on Vera Rubin and agent harnesses run on Vera CPUs. BlueField-4 units run Nvidia Sentry, which watches agents from outside the main system and quarantines any that leave their boundary within milliseconds.³

The design builds on Nvidia's open source OpenShell runtime, released under Apache 2.0. It extends to RTX Spark, DGX Spark and DGX Station machines.³ Nvidia says agents from Anthropic, Google, Meta, OpenAI and others can run on it, and security vendors including CrowdStrike, Okta and Palo Alto Networks are adding support.³

ServiceNow made its long-running Arc agent generally available with actions enforced by OpenShell. Dell, Cisco, HPE, Lenovo and Supermicro are building infrastructure on the design.¹⁹ Nvidia did not give prices.³

Why it matters for builders: Enforcing limits outside the agent, in the runtime and in separate hardware, means a compromised or confused agent cannot turn off its own guardrails.⁴ Teams running agents with real access should plan for that kind of design.¹⁴

In brief

Funding & deals
OpenAI models will also reach Google's Antigravity coding tool and Gemini Agent. Beyond Model Garden, the deal brings OpenAI models to Antigravity and Gemini Agent, and Google Cloud CEO Thomas Kurian framed it as letting customers pick the right model per job.²⁰ Antigravity already offers Anthropic's Claude models to its users.²⁰
Industry
Dario Amodei defended Anthropic's safety record in a long reply to a critic. Responding to Encode AI's Nathan Calvin, Amodei listed nine decisions he said cost Anthropic commercially. One was holding Mythos-class models back from general release from February until June, until stronger cyber safeguards were in place.²¹ He conceded that some policy communication, including a draft bill sent to Congress, could have been handled better.²²
Infrastructure
Nvidia's GTC Berlin keynote also covered robotics tools and European partnerships. Nvidia shipped a new Isaac simulator with expanded APIs and template workflows for its GR00T robot model. It also showed Agile ONE SN, a reference humanoid robot from Agile Robotics.²³ Huang announced autonomous vehicle work with Mercedes-Benz and VW, and Deutsche Telekom plans to expand its Nvidia-powered Industrial AI Cloud.^{23, 24}
Products
Walmart is giving outside AI shopping agents their own profiles and spending limits. Walmart introduced Agent Profiles, the Walmart Agent Gateway and new terms for AI agents, letting customers shop through the assistants they choose with controls built in.^{25, 26} Profiles arrive within a month and let shoppers require approval for every purchase or set a cap such as $50, $100 or $200. Amazon, by contrast, has blocked Meta's Muse.^{6, 25}
Models
Mistral released Voxtral 3, a family of speech models with a voice agent SDK. The release includes three transcription models, a text-to-speech model priced at $13 per million characters, and a LiveKit-based SDK that chains listening, reasoning and speaking into one pipeline.²⁷ Batch transcription starts at $0.0035 per minute. The realtime model is less accurate than Gemini 3.5 Realtime, and Mistral announced no open weights.²⁷
Products
Meta says Muse has passed 10 million users, with shopping among the top uses. On Meta's third-quarter call, Mark Zuckerberg said about 25% of people who set up Muse use it daily. He added that over a quarter of engaged users shop through it, with average purchases above $200.²⁸ Meta expects to cut Muse's compute costs by more than 10x next year and is planning a Muse network where agents deal directly with businesses' agents.^{28, 29}
Dev tools
AMD and Intel are merging their CPU inference optimizations for PyTorch into oneDNN. AMD is folding its ZenDNN library into the open source oneDNN and joining the UXL Foundation as a steering member. The result is a single PyTorch backend tuned for both companies' chips.³⁰
Funding & deals
Red Hat and Nvidia expanded their partnership on agent security and model governance. Red Hat is working to make Nvidia's OpenShell Reference Architecture a native capability of its AI platforms. It also published a joint reference design for model access controls built on vLLM, llm-d and Nvidia Dynamo.⁴ Red Hat Enterprise Linux for Nvidia now supports Vera CPUs.⁴
Products
ChatGPT's Intelligent UI can build playable games inside a conversation. The GPT-6 feature lets responses include buttons, forms, charts and interactive experiences. A TechRadar test produced working versions of Snake, Tetris and Pac-Man, though the model sometimes returned code instead of a playable result.³¹
Infrastructure
Anthropic reportedly plans to spend $750 billion on compute in the coming years. The figure, which includes hundreds of billions in non-cancelable deals, comes from a prospectus reported by The Wall Street Journal ahead of an IPO that could value Anthropic at $2 trillion.³²
Industry
Cloudflare says more than 10 million developers now build on its platform. That is up from 7.4 million a quarter earlier, and Cloudflare ties the growth to AI agents.^{33, 34} Daily requests from AI agents on its network grew more than 1,700% in a year, though the company did not say how many of the developers pay.³³
Funding & deals
Flapping Airplanes raised a $1 billion Series A at a reported $6 billion valuation. The year-old lab, co-led by Index and Lux in this round, is building a megakernel framework to train models with far less data and keep GPUs from sitting idle.³⁵
Research
UCSB and Microsoft researchers built ScholarEvolve to upgrade agent harnesses using techniques from research papers. The framework finds failures, retrieves relevant arXiv papers and turns their techniques into code for the harness. It raised Qwen3.5-27B from 69.0% to 81.4% on AppWorld without changing the model's weights.³⁶
Products
Google Public Sector launched a Gemini agent product for US government services. It handles policy analysis, Freedom of Information Act redactions and procurement compliance, and Missouri and the Department of Health and Human Services are already using it.³⁷
Products
Epic is exploring clinical AI agents that could handle routine visits. Epic executives discussed the idea at a fireside chat in Harvard Medical School's Clinical Informatics Lecture Series.³⁸

That’s the edition.

The next one is ready tomorrow by 6 AM ET.

Wed, Oct 21 edition All editions

Written with AI from the sources in the references above.

AuthorRamit KoulFounder, Software Engineer & Innovator