What jumped out at me is not that Anthropic built its own version of Meta Muse or OpenAI Dots. That was inevitable. The interesting part is how careful and slightly defensive this launch sounds. Anthropic is basically saying: yes, the agent can wander through your connected accounts, but only a little, only sometimes, and only after we’ve put several guardrails around it. That tells me they know exactly how uneasy people are going to be about this category.
The most promising detail is also the least flashy one: Duet runs on its own virtual machine and doesn’t need a Claude session open the whole time. That matters. If you’re going to sell “proactive” AI, it has to live somewhere other than a chat window, or it’s just a nicer prompt template. A background worker with a UI is much closer to a real agent than the usual demo theater.
But I’m not sure I buy the “short briefing with action items from your connected apps” pitch yet. The moment an agent starts checking accounts “every few hours,” the tradeoff gets real. Either it’s too timid and misses things, or it gets noisy and burns through usage. Anthropic seems aware of that problem, which is why the article mentions tuning message frequency and topic controls. Good. Also kind of revealing. If the core experience needs that much user calibration, maybe the product is still closer to an opinionated notifier than a trusted assistant.
The pricing angle is the part I’d actually watch. If Pro subscribers eventually get access, that’s a much more approachable lane than the ultra-premium setup other companies are using. I think Anthropic is trying to position Duet as something practical instead of status-y: less “look at our futuristic personal agent” and more “here’s a thing that might save you some admin work if you trust it.” That’s smart, because trust is the whole game here, not model quality alone.
And trust is where the article lands hardest, probably unintentionally. The security language reads like the company already knows the obvious objection: if an AI is touching your life, your calendar, your work tools, maybe even finance-adjacent data, then the scary part is not hallucination, it’s exposure. The claim that passwords and tokens never reach Duet is reassuring in the abstract, but I’d want to know much more about the actual permission boundaries, failure modes, and auditability before I let it do anything nontrivial. “We have a classifier” is not the same as “I’m comfortable with this.”
So my reaction is mixed. Duet sounds more grounded than the usual agent launch, and that’s a good sign. But the whole category still feels like it’s being sold a year or two before most people will really know what they want from it. I’d try it, sure. I just wouldn’t hand it the keys to anything important on day one.
Reference: Claude Duet is Anthropic's Answer To Meta Muse And OpenAI Dots
