Always-On AI Agents Are The Race; The Switchboard Is The Contest

🔍 Read the full analysis: Always-On AI Agents Are The Race; The Switchboard Is The Contest on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Anthropic launched Claude Duet on October 21, adding a background-working AI agent to a fast-growing field that includes products from OpenAI, Meta and SpaceXAI. The waitlisted research preview is available to US Pro and Max subscribers and is designed to work from a cloud computer across connected apps, while Anthropic says users must approve sensitive actions such as sending messages or making payments.

Duet is a separate app available through iOS, Android and claude.ai, according to Anthropic. It can provide a morning brief, accept tasks by text or voice, and send a notification when it needs a person to make a decision. Its work runs on a cloud-based virtual machine, so the user’s own computer or phone does not need to stay on. Anthropic says the preview uses Claude Opus 5.5 and carries over memory, skills and connected apps from a user’s Claude account.

Anthropic says Duet can read information and prepare work, including scheduling, dashboards, sales forecasts and small web apps. It cites tasks involving QuickBooks, Stripe, Gusto, DocuSign and Shopify, and says Claude Marketplace offers more than 6,000 connectors and plugins. Those descriptions are company claims; the source material reports no independent testing of the preview’s capabilities or safeguards.

In a safety post dated October 23, Anthropic said Duet displays the action, destination and content of a sensitive message or payment and waits for approval before acting. The company says trading securities, moving money between accounts and accessing another person’s account are blocked. It also says certain actions, including paying, sending money and deleting, cannot be set to run without repeated approval. For purchases, Anthropic describes a one-time Stripe Link card limited to a merchant and amount. These controls and restrictions have not been independently verified.

Control of Apps May Define the Agent Race

The launch points to a competition over more than which company has the strongest model. Agents that work across email, calendars, payments and business software need permission to access those systems, a continuing record of user preferences and a way to decide which model handles each task. Whoever controls that layer can shape what the agent sees, what it can do and which services receive the work.

The source material describes SpaceXAI’s Grok Bot as able to route tasks to external models, including Claude Opus 5.5. It also cites xAI’s security FAQ as saying Grok Bot’s model choice cannot be restricted through a team model allowlist and that there is no customer-facing model picker. If that description holds, users may interact with one assistant while their requests are handled by different providers. That makes routing, account access and usage costs meaningful questions for customers, not just technical details.

For businesses, persistent access can make routine work easier to delegate, but it also creates risks around sensitive data, mistaken actions and staff approving prompts without careful review. Anthropic itself warns about approval fatigue: frequent confirmation requests can train people to click through them. Companies testing agents will need to judge both their productivity and how well permission controls work in practice.

A Cluster of Agent Launches

The product arrived amid a sequence of announcements described in the source material: SpaceXAI opened a Grok Bot beta on August 11, Meta launched Muse on September 8, OpenAI introduced Dots at its September 29 DevDay event, and Google released Gemini Agent on October 8, according to TechTarget. Anthropic announced Duet on October 21. The products differ, and the supplied material does not provide a full, independently verified comparison of their features or performance.

The shared product idea is an assistant that can act beyond a single chat session: it keeps working in a hosted environment, uses connected services and returns with completed work or a request for input. Duet’s early-access limits are material: it is a research preview by waitlist for US Pro and Max subscribers, not a generally available service. Its real-world reliability and usefulness remain to be established.

“for professionals and small business owners with more on their plates than they can track.”

— Anthropic, describing Claude Duet

Preview Limits and Safeguards Remain Untested

Anthropic says Duet can misread message threads or misjudge urgency. It also says the preview is weaker at personal errands such as booking tickets, cannot yet show users its browser and checks connected accounts every few hours rather than continuously. Those checks count toward the same usage limits as other Claude activity, and large tasks may require more model turns.

The available information does not establish Duet’s accuracy, security or effectiveness in independent tests. Anthropic says the agent does not see passwords, browser cookies or sign-in tokens and that a separate classifier checks whether proposed actions match user intent, but the supplied material contains no outside validation of those claims. The company also says disconnecting an app prevents future access, while information already learned remains until the user resets Duet’s memory. How users can inspect or manage that retained information beyond a reset is not detailed here.

It is also unclear how Duet’s usage limits will affect regular workloads, how often it will request approval, and how its safety controls compare with those of rival agents. The source material does not provide confirmed pricing for Duet beyond its eligibility for Pro and Max subscribers, or results from a public performance evaluation.

Access and Testing Will Shape Adoption

Anthropic has not provided a date in the source material for ending the waitlist or making Duet generally available. The next useful evidence will come from broader access, clearer product terms and independent testing of the agent’s performance, app permissions, approval process and memory controls.

For now, the source recommends that companies run small, written pilots with limited, read-mostly permissions and a defined exit plan. Users can begin with low-stakes tasks, connect only a few apps and keep approval requirements enabled for actions that send, pay or delete. Those precautions do not establish that the product is safe; they limit exposure while its behavior is being evaluated.

Further launches and product changes from Anthropic and its competitors will show whether these agents can reliably move from demonstrations to routine work. For customers, the practical test will include not only whether an agent completes tasks, but also who can access its data, choose its models and control its actions.

Key Questions

What is Claude Duet?

Claude Duet is Anthropic’s separate AI-agent app, introduced as a waitlisted research preview. Anthropic says it can work in the background on a cloud computer using connected apps and return with a brief, completed task or request for input.

Who can use Claude Duet?

According to the source material, access is limited to US subscribers on Claude Pro and Max plans who join the waitlist. It is available through iOS, Android and claude.ai.

Can Duet send messages or make payments by itself?

Anthropic says Duet must show the details and get approval before sensitive actions such as sending a message or making a payment. The company says some actions, including trading securities and transferring money between accounts, are blocked even with user approval. These statements have not been independently tested in the supplied reporting.

What are the main reported limitations?

Anthropic says Duet may misread threads or urgency, is less capable at personal errands such as booking tickets, cannot yet show its browser and checks connected accounts every few hours. Background checks and large tasks also use the plan’s shared usage limits.

What should organizations test before adopting it?

The source recommends a small pilot with written boundaries, limited permissions, a read-mostly starting point and an exit plan. Teams should assess reliability, approval frequency, data access, memory retention and any model-routing controls before using an agent for consequential work.

Source: ThorstenMeyerAI.com