In one minute
- Anthropic introduces Claude Duet: it works in the background, but access is restricted at launch and Microsoft 365 is not yet available.
- Walmart prepares for shopping by agents: customers set limits and can withdraw permission; NVIDIA proposes external controls with OpenShell.
- OpenAI comes to Google Cloud in a private preview: invited businesses will be able to test its models under Google’s administration.
- Volkswagen expands Copilot and Mistral proposes new tests: more agents in businesses means a greater need to know what they do and whether they do it well.
Lead story
Anthropic introduces Claude Duet, an assistant that prepares work on its own initiative
Agents and work · Announced: 21 October 2026
What it does
Claude Duet is a new app based on Claude Opus 5.5. It can review connected apps, prepare a morning brief, arrange meetings, identify unpaid invoices and produce reports or financial forecasts. It runs in a virtual machine with its own browser. It can keep working even when the user turns off their devices. It checks accounts at intervals, every few hours. [1]
Who can use it
It starts as a research preview, with access through a waitlist and invitation, for Pro and Max users in the United States. It has iOS and Android apps and web access. It is not yet available for Team or Enterprise, or to users in Spain or Norway. Microsoft 365 is not available either. Background work counts towards the same plan allowance as conversations with Claude. [1][2]
What requires approval
It is designed to show the action and wait for approval before sending messages, making payments, signing documents or deleting information. Trading securities or moving money between accounts remains prohibited even if the user approves. It also cannot access files on the user’s computer; the user cannot watch its browser or take control of it. [2]
Why it matters: the new element is an assistant that tries to decide when work needs doing. For a law firm, reading a case file, preparing a response and sending it require different permissions. Those permissions should be granted separately and be revocable.
Verification limits: these are features and safeguards described by Anthropic. The company acknowledges that Duet may misread a conversation or misjudge urgency. Human approval remains a decision that requires review. [1][2]
Sources: 1 · Anthropic, announcement · 2 · Official help: access, permissions and limits.
Other news
2. Walmart prepares profiles and spending limits so agents can shop for customers
Commerce and permissions · Announced: 21 October 2026
Walmart introduced Agent Profiles, linked to the customer’s account, which will make it possible to distinguish each agent’s purchases, set limits and revoke permissions. Purchases will require approval by default. Customers will be able to allow transactions up to a set amount; Walmart may request additional confirmation. Agents will not receive card details or be able to change payment methods. [3]
Why it matters: permission to buy will have a defined scope and can be withdrawn. That makes the delegation of a task involving financial commitments concrete.
Verification limits: the profiles are scheduled for November. Walmart Agent Gateway, the connection for external agents, starts in an invitation-only beta. These are announced terms, not general availability. [3][4]
Sources: 3 · Walmart, announcement · 4 · Walmart Agent Gateway.
3. NVIDIA introduces OpenShell to monitor agents from outside their environment
Agent oversight · Announced: 21 October 2026
NVIDIA introduced OpenShell Reference Architecture, an architecture that separates the execution of agents from their supervision and the enforcement of rules. It proposes external controls that can isolate and stop an agent if it tries to move beyond its permitted boundaries. Work and oversight therefore do not depend on the same component. [5]
Why it matters: granting permissions achieves little if the agent itself can change the mechanism that enforces them. When procuring an agent, it matters who sets its limits and who can stop it.
Verification limits: NVIDIA says the architecture is available. Its isolation capabilities are the manufacturer’s claims; a technical architecture does not, by itself, certify legal compliance. [5]
Source: 5 · NVIDIA, OpenShell Reference Architecture.
4. OpenAI and Google Cloud open a new route for businesses to use OpenAI models
Business and cloud · Announced: 21 October 2026
Google Cloud is adding GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna to Model Garden, its model catalogue. It also allows local Codex workflows to connect to its infrastructure. The offering is purchased through Google Cloud Marketplace and can count towards existing spending commitments. Google announces unified identity management and audit logs. [6][7]
Why it matters: bringing models under the same procurement and administration arrangements may make business adoption easier. The controls covered by each service will need to be checked.
Verification limits: initial access is private, by invitation, for development and internal testing. OpenAI expects general availability in the coming weeks. Zero data retention and Confidential Compute options remain planned for a later date. [6][7][8]
Sources: 6 · Google Cloud · 7 · OpenAI, announcement · 8 · Initial access terms.
5. Microsoft puts Volkswagen’s Copilot rollout at more than 330,000 employees
Business adoption · Announced: 21 October 2026
Microsoft reports that Volkswagen has equipped more than 330,000 employees with Copilot. It adds that more than 26,000 agents have been active across the group in the past month. The rollout covers engineering, production, procurement and sales tasks. These are two different measures: employees with access and agents in operation. [9]
Why it matters: when agents number in the thousands, oversight needs accountable people, permissions and logs. It is no longer enough for each employee to know the precautions that apply to their assistant.
Verification limits: the figures come from Microsoft. The announcement does not detail the controls on those agents or provide an independent productivity assessment. [9]
Source: 9 · Microsoft, Volkswagen rollout.
6. Mistral and NVIDIA prepare public tests for engineering agents
Evaluating results · Announced: 21 October 2026
Mistral and NVIDIA announced a collaboration to apply AI to design and simulation. They will prepare public tests and open data to evaluate models of physical systems, alongside specific evaluations of engineering agents. These will measure tool use, execution across multiple steps, technical correctness, time and cost. [10]
Why it matters: an agent can stay within its permissions and still deliver the wrong result. Before assigning it specialist work, it helps to be able to compare its results using tests others can repeat.
Verification limits: the announcement describes a development programme. It does not present a complete suite of tests that has already been published and validated. [10]
The signal in the noise
Agents are receiving authority to act. The important decision is who grants it, who limits it and who monitors its use. A law firm can start with something as simple as preparing replies without sending them. A business can authorise purchases up to a set amount. In both cases, responsibility and the way to withdraw permission must be clear.
Duet and Walmart illustrate delegation. OpenShell proposes enforcing its limits. Mistral adds verification of the result. That progression matters more than the promise of an assistant that does everything.
On the radar
Spain · A law firm’s advantage lies in how it integrates AI
Cinco Días reports on an Aranzadi LA LEY discussion about the competitive advantage of integrating AI well into a law firm. LegalToday, the organiser’s portal, confirms the event. These are participants’ views, not an independent measure of productivity. Both reports were published on 21 October; Cinco Días identifies its piece as a collaboration with LA LEY. [11][12]
Sources: 11 · Cinco Días · 12 · LegalToday / Aranzadi LA LEY.
Norway · Digital inclusion and the AI Act are two separate matters
On 21 October, Digdir published the results of its 2023–2025 digital inclusion plan: 59 of 71 measures have been completed or are continuing. This is a review of digitalisation in general. Separately, the EFTA records consulted on 22 October still list the AI Act and the Digital Omnibus on AI as under consideration for incorporation into the EEA Agreement. The EU timetable alone does not establish that they are in force in Norway. [13][14][15]
Sources: 13 · Digdir, final report · 14 · EFTA, AI Act · 15 · EFTA, Digital Omnibus on AI.
Biology · Agents to identify what to investigate next
On 21 October, OpenAI, PerturbAI and Scripps Research reported an analysis using thousands of GPT-Rosalind agents on mouse brain cell data. The result is a set of hypotheses to guide subsequent experiments, not proven treatments. [16]
Source: 16 · OpenAI, agents and biological data.
Infrastructure · NVIDIA extends protection for models and data
On 21 October, NVIDIA introduced Confidential Bare Metal, an option to protect models and data while they are processed on dedicated third-party servers. It adds another layer of control: over access to information. The announced integrations have different timelines. [17]
Source: 17 · NVIDIA, confidential computing.