Anthropic Joins the Always-On Agent Race With Claude Duet

Anthropic just dropped its answer to the "always-on" AI agent race, and the timing says as much as the product itself. Claude Duet, announced on October 21 as a research preview, is the third major persistent agent to launch from a top US AI lab in under two months. That is not a coincidence. It is a category forming in real time.

The concept is simple on paper: an AI that does not wait for you to type a prompt. It runs in the background, monitors your connected apps, surfaces what needs attention, and keeps working after you close the app. Meta got there first with Muse in late August. OpenAI followed with Dots at DevDay in late September. Now Anthropic is joining with a product built around Claude Opus 5.5, the same model the company recently positioned against GPT-6 Astra on coding benchmarks.

But Anthropic is not just copying the playbook. It is making a bet on restraint.

The Approval Gate as Product Strategy

Where rivals have emphasized speed and autonomy, Anthropic is building in checkpoints. Claude Duet is explicitly designed to ask for permission before sending a message on your behalf, making a payment, or deleting anything. The company frames this as a feature, not a limitation. In a category where user trust is still the biggest open question, Anthropic is essentially saying: the agent that asks before it acts is the one you will actually let into your workflow.

This is consistent with the company's broader positioning. Anthropic has spent years talking about AI safety, constitutional AI, and interpretability. A cautious agent is the natural product expression of that philosophy. Whether the market rewards that caution, or whether users gravitate toward the more autonomous alternatives from Meta and OpenAI, is the experiment Duet is designed to answer.

What Duet Actually Does

The practical pitch is straightforward. Each morning, Duet sends a brief summarizing the most important items across your connected apps. During the day, it messages you only when something needs attention. You can also hand it tasks ranging from simple scheduling to more complex work like building dashboards, creating forecasts, or prototyping web apps.

Behind the scenes, Duet runs on a virtual machine with a browser, so it can complete tasks that require web interaction. It pulls in memory, skills, and connected apps from the main Claude experience. For sensitive actions, it has a new "dedicated classifier and safety system prompt" designed to reject requests to share the strictest tier of information, such as credit card numbers.

Anthropic is candid about the limitations. Duet is an "early version" that "still might misread a thread or miscalibrate what is urgent." It is also not as strong at personal errands like booking tickets. The company is launching it as a research preview through a waitlist, and only for Claude Pro and Max subscribers in the United States.

The Competitive Landscape

The timing of these three launches is striking. Meta shipped Muse first, leveraging its massive distribution through WhatsApp, Instagram, and Messenger. OpenAI followed with Dots, built to integrate across apps and scale through the company's existing developer and consumer relationships. Google's Gemini has had agentic capabilities too, though it has been quieter about the consumer rollout.

Anthropic is entering last, with a smaller consumer footprint than Meta or OpenAI. But it is not without advantages. Claude has built a reputation as the most reliable model for complex, long-form tasks. Opus 5.5 has been well received by developers who need depth over speed. If Duet can translate that technical credibility into a trustworthy agent experience, Anthropic could carve out a meaningful position even without the distribution scale of its rivals.

What This Means for the Category

The emergence of three persistent agents from three major labs in under two months suggests we are past the experimentation phase. This is now a strategic front in the AI platform wars. Each company is making different bets about how much autonomy users want, how much oversight they need, and where the line between helpful assistant and invasive surveillance sits.

Meta is betting on ubiquity: get Muse into as many hands as possible through existing apps. OpenAI is betting on integration: make Dots work across the broadest set of tools and services. Anthropic is betting on trust: build the agent that users feel safest delegating to, even if that means moving slower.

The question now is which of these bets pays off. Users have not yet voted with their attention, let alone their wallets. Most of these products are still in limited previews or bundled into existing subscriptions. The real test will come when they are unleashed at scale, when the edge cases pile up, and when someone inevitably gets burned by an agent that acted too fast or asked too often.

The Bigger Picture

Beyond the consumer race, persistent agents raise broader questions about labor, attention, and accountability. If an agent schedules a meeting incorrectly, who is responsible? If it misses an urgent message because its classifier judged it non-critical, what is the recourse? These are not hypotheticals. They are the design questions that will shape how these products evolve.

Anthropic's approach of building explicit approval gates is a recognition of this reality. Whether users find those gates reassuring or annoying will determine a lot about which agent becomes the default. The company is essentially running a live experiment in human-AI collaboration, with the hypothesis that trust, once earned, scales faster than features.

For now, Claude Duet is a research preview available only to paying subscribers in the US. But its launch confirms what the last two months have made increasingly clear: the age of the always-on agent is here, and the competition to define it is just getting started.