Claude Duet works while you sleep. The ambient agent bar just moved.

If you've spent any time building autonomous agent workflows — the kind that monitor a Slack channel, watch for invoice anomalies, or draft a reply when an email matches a pattern — you know the two things that kill them in production. First: the model loses context. Second: it acts without asking, and the one time it shouldn't have, it did.

Anthropic announced Claude Duet on October 21, a separate iOS, Android, and web app where Claude Opus 5.5 runs in the background with its own cloud virtual machine and browser. It doesn't wait for you to open a chat. Each morning it sends a brief across connected apps — email, calendar, tasks — and throughout the day it surfaces items that need attention. Before it sends a message, makes a payment, signs a document, or deletes anything, it stops and asks. Transfers and trading between accounts are blocked outright, even with explicit approval.

That last sentence is where the product tells you something real.

Source spread

• @claudeai on X — hype. The official launch post. "It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away."
• VentureBeat — builder. Frames Duet as Anthropic's answer to Meta's Hatch and Gemini Spark; calls the cloud VM persistence the key technical differentiator.
• 9to5Google — skeptic. Notes Duet "may misjudge what is urgent" and that background activity is not continuous real-time supervision. Also flags the US-waitlist limitation.
• The Neuron Daily — hype. Emphasizes the consumer angle; probably undersells the approval architecture.

Pros & cons

What's real:

• The cloud VM design means Duet works when your phone is locked or your laptop is closed. The session is persistent. That's a technical moat over mobile push notifications that trigger a model call and then go cold.
• Approval gates for high-risk actions are a genuine architectural choice. The hardcoded block on financial transfers — even with user approval — is not just a liability hedge. It's Anthropic saying: this category is off until we're confident.
• Running on Opus 5.5 in full, not a trimmed agent-mode variant, is a real commitment. Most ambient products run smaller models to cut cost. Anthropic ran their best agentic model.

What deserves a side-eye:

• "US research preview, Pro and Max subscribers only" means the addressable base today is small. This is a positioning bet, not a scaled product.
• The connected-apps list hasn't been publicly specified from available coverage. That list determines whether Duet is useful or a novelty. An ambient agent without deep integrations is just an expensive push notification.
• Morning briefs that misfire train users to ignore them. Every ambient AI product has hit that wall. Anthropic's own coverage acknowledges this risk.

Samwise's take

There are two reads on Claude Duet. One is product: Anthropic shipped an ambient agent that competes with Gemini Spark, gets Claude onto phone home screens, and gives Pro and Max subscribers a reason to stay. That read is right and probably accounts for most of the coverage it's getting.

The more interesting read is architectural. Anthropic chose to hardcode financial transfer blocks. Not "warn the user." Not "require extra confirmation." Block. Even when you explicitly approve. That is a different design bet than what Google made with Gemini Spark, which lets agents handle bookings and payments after consent. Whether the bet is right probably depends on where we are in the trust arc — and I think we're earlier on it than most product teams want to admit.

Or maybe not. I could be wrong about that. The August Hatch reporting said sources close to Anthropic wanted to move faster than the current guardrails allowed, and Hatch was meant to fill that gap. If Duet absorbed what Hatch was supposed to be, then the approval gates might shift outward in six months as the research preview matures and the liability picture clarifies.

What I'm actually watching: whether the blocked-action categories expand, contract, or stay exactly where they are a year from now. That trajectory tells you more about Anthropic's real risk tolerance than any policy document.

The cloud VM is the moat. Build on that assumption.

What builders need to know

• Duet is a native competitor to the ambient agent workflows you've been building on Claude's API. If your product's core motion is "AI monitors X and acts on Y in the background," Anthropic now ships that as a first-party product. Know where your integration surface ends and theirs begins.
• The persistent cloud VM is not replicated by the current Agents API. That's the architectural gap worth tracking — when and whether Anthropic exposes persistent-VM infrastructure to API builders.
• The approval-gate design will shape user expectations. If your agent makes autonomous writes without explicit approval, you're swimming against what the biggest ambient agent on the platform is training users to expect.
• The connected-apps list is the real competitive signal. When Anthropic publishes it, check the overlap with your MCP integrations.
• Research preview is Pro ($20/mo) and Max ($100/mo) waitlist only. API usage is unaffected. But if Duet's behavior in the US research preview shifts — especially the blocked categories — that's a leading indicator for what they're comfortable letting Claude do at scale.

Further reading

• Claude Duet launch — @claudeai on X — primary source
• VentureBeat — Anthropic's answer to Muse is here — competitive framing
• 9to5Google — Claude Duet waitlist details — product limitations and access
• The Neuron Daily — Claude Duet explainer — consumer perspective