Claude Duet: Anthropic's always-on agent runs on Opus 5.5
Two launches in two days, and they are about the same thing: an agent that acts for you needs an app first, and then it needs your passwords.
Anthropic opened Claude Duet to a US waitlist on Wednesday — a standalone agent app for iOS, Android and the web that runs on Opus 5.5, works on its own cloud computer, and messages you when something needs a decision. It ships inside Claude Pro and Max rather than as a separate subscription, carries over your existing memory, skills and connectors from the Claude app, and reaches more than 6,000 apps and plugins, according to Anthropic's Claude Duet page. Team and Enterprise plans are not included yet.
What separates Duet from the chat app is the direction of travel. Claude waits for you to open a thread; Duet comes to you — a morning brief of what is slipping, drafted next steps, then quiet until something needs a human. The examples Anthropic puts on its own product page are small-business operations rather than demos: a county health permit fee waiting on a $640 approval, a renewal redlined to Net 60 with the changes already summarised, 27 wrong delivery orders with refunds drafted. Tasks arrive as messages — "build me one page I can check every morning" — and come back finished.
The permission model is the part worth reading closely. Duet asks before sensitive operations like sending messages or paying, and the approvals, logins and safety checks run inside your own cloud computer, so the sensitive details stay out of the conversation. That is the right architecture, and it is also a claim about something no lab has demonstrated at scale yet: that an agent holding standing access to your inbox, calendar and payments asks at the right moments, and not at all the others.
Anthropic is unusually direct about the gaps. Its own FAQ says Duet "may misread a thread or miscalibrate what's urgent, and it can miss something that arrives between checking in with you." It also shares your plan's usage limits with the rest of Claude, and background checks draw on those limits — a quiet way of saying an always-on agent burns a subscription faster than chat does. There is no Team or Enterprise tier, no published retention, and no date for leaving research preview.
The competitive frame is unavoidable. OpenAI's Dots shipped at DevDay on the identical pitch — an agent with its own cloud computer, more than 4,000 connected apps, "proactive research" — and we covered what that launch promised in OpenAI's Dots give Pro users an agent with its own computer. Meta's Muse got there first in September and still has the only public usage numbers: more than 10 million users, over a quarter of them buying, at an average order above $200, per Mark Zuckerberg on the earnings call. Anthropic's differentiator is not a longer feature list. It is Opus 5.5 — the model buyers already trade down to on price — bundled where Pro subscribers already sit.
1Password launched 1Password for Agents today, a sign-in layer that lets an AI agent log into your accounts without the password ever reaching the agent. The mechanism is deliberately boring, which is why it might work: the agent presents a unique digital key when it reaches a login form, 1Password tells you which account it wants and why, you approve on your own device — the same gesture as a two-factor prompt — and 1Password fills the credentials itself. Agents from Meta, OpenAI, Anthropic and SpaceXSI are supported, SiliconANGLE reports, with ChatGPT Work, Dots and Muse live today and nine more integrations in progress, including Perplexity and Browserbase.
The Claude Duet line matters more than the launch list. Anthropic chief product officer Ami Vora told SiliconANGLE that 1Password will "soon" fill in Duet's sign-ins using passwords people have already set up, aimed at users "who've never before let an agent near their accounts." Duet already asks before signing in anywhere; what was missing was where the password came from. Pasting it into a chat window, which is what most people do now, leaves a copy outside anyone's control — and 1Password is not saying when each partner integration reaches users.
Aashish Tripathi of 1Password framed the launch as a break with that habit. "The old model where you hand over a credential and hope for the best doesn't work," he said. The design also commits the company to the position its chief technology officer, Nancy Wang, argued at Oktane last month: access tied to a single task, where an agent has to prove it finished the last one before it gets more. The risk nobody has priced is approval fatigue. A security prompt that arrives often enough becomes a tap, and a tapped prompt is not consent.
Meta's quarterly filing puts its total future commitments for leases and contracts at about $825 billion, up from nearly $700 billion in July, with off-balance-sheet lease obligations of $355.7 billion running to 2036, as Bloomberg reported. The disclosure landed beside results that were strong at the top and thin underneath: revenue of $64.6 billion, up 26% year over year, operating income down 37% to $12.99 billion after a $9.89 billion legal settlement charge, and family daily actives at 3.64 billion.
The number to hold onto is the gap between $825 billion of commitments and a business earning roughly $13 billion a quarter from operations. Commitments are not debt, and long leases on datacentres are exactly what a company signs when it expects a decade of demand. But they are also a bet that somebody pays for the inference — and the only agent in this story with public revenue numbers is Muse, at 10 million users and a shopping habit. Meta is separately creating a class of nonvoting stock, effective January 2027, to raise capital for the buildout without diluting founder control.
Put the three stories together and the shape of the next year is visible: the agents shipped, the credential layer arrived, and the bill was signed. What none of it settles is whether the demand underneath — people handing an always-on program their inbox, their calendar and their card — is as large as the commitments assume.
What to watch: whether Duet's waitlist opens beyond US Pro and Max before the year ends, and whether Anthropic publishes usage limits specific to the agent rather than sharing the plan's pool.
If keeping an agent safe means approving forty prompts a day, will you still be reading them by Friday? Write to us via our contact page.
