An agent that knocks
The pitch for Claude Duet, in Anthropic's own words, is a "proactive agent that understands what needs your attention." You hand it work through messages, it works on its own computer in the cloud using the apps you connect, and it messages you back when it's done — or when it needs a decision. It launched in October 2026 as a research preview, on a waitlist, for Claude Pro and Max subscribers in the US. Team and Enterprise plans are not included yet.
The design is a deliberate inversion of the chatbot relationship. Instead of you doing the work with an AI assistant looking over your shoulder, Duet does the work on its own machine and you supervise at the moments that matter. The approval gate is the load-bearing feature: Duet requires your approval before it sends messages, makes payments, or edits documents. Anthropic is selling autonomy, but gated autonomy — the agent acts freely until the action is consequential, at which point it knocks.
Under the hood, Duet runs on Opus 5.5 and plugs into more than 6,000 apps and services. The connector count matters because a proactive agent is only as good as the tools it can reach. An agent that can see your calendar, read your docs, and act in your apps is a fundamentally different creature from one that just talks.
The handoff model has a specific failure mode Anthropic is trying to avoid: the agent that does nothing until asked. A purely reactive assistant leaves the user as the bottleneck — every task still starts with a human deciding to delegate it. Duet's proactive framing is meant to close that loop: the agent is supposed to notice the expense report sitting in your inbox, or the calendar conflict forming next week, and start working before you think to ask. Whether that feels magical or invasive will depend entirely on the quality of the judgment calls about "what needs your attention," which is presumably why the preview is gated to individual power users first.
There's also a competitive logic to the connector-first strategy. Every AI lab is racing to ship agents, but an agent without access to your tools is a demo, not a product. By launching Duet with 6,000+ connectors on day one, Anthropic is arguing that the moat in proactive agents isn't the model — it's the wiring into the stack where your work already lives. The agent that can act inside your existing apps beats the agent with the cleverer conversation.
The launch reaction
The launch didn't happen quietly. Anthropic's official @claudeai announcement on X on October 21 was met with the full spectrum of reactions — excitement, skepticism, and at least one now-famous one-liner. 9to5mac's coverage embedded the launch announcement alongside CZ's reaction: "Now, it controls us." The joke landed because it names the real anxiety: every step toward proactive agents is a step toward software that acts while you're not looking.
Anthropic's answer to the anxiety is architectural rather than rhetorical. The approval-before-action design, the research-preview framing, the US-only waitlist for individual Pro and Max subscribers — this is a company releasing a capable agent while visibly keeping its hands on the guardrails. The forum thread on Product Hunt about the Duet launch picked up the same tension: builders are excited about an agent that handles real workflows, and wary about what "proactive" means when the machine decides what needs your attention.
Why the company behind it matters
Duet arrives from a company that has been unusually transparent about its thinking on AI agency. Dario Amodei, Anthropic's CEO and co-founder (with his sister Daniela Amodei), has spent years in long-form interviews — including a well-known appearance on the Lex Fridman podcast (episode #452) — laying out the case that increasingly capable AI systems are coming, and that the industry's job is to steer them. Duet is the productized version of that argument: an agent powerful enough to do real work, with explicit human checkpoints.
The company behind Duet is not a scrappy startup testing a thesis — it's the best-funded independent AI lab in the world. May 2026 coverage (tech-insider.org) reported a $65 billion Series H at a $965 billion valuation. That scale is relevant to the product: a proactive agent that touches your payments and your documents requires a level of trust, infrastructure, and safety work that only companies with Anthropic's resources can fund. The waitlist, the preview label, the US-only rollout — read them as a company that can afford to be careful.
Which makes Duet a useful lens on the agent race. The question Anthropic is betting on isn't whether agents will do useful work — it's whether the winning agent is the one you talk to, or the one you hand things off to. Duet's answer is the handoff: delegate, get pinged when it matters, approve the consequential moves. If that interaction model sticks, the chatbot era will look, in retrospect, like training wheels.
The Fridman interview is worth revisiting in light of Duet because it shows the continuity of the thinking. Amodei has argued for years that the industry should expect AI systems to take on progressively more autonomous roles, and that the responsible move is to build the steering mechanisms — like approval gates and research previews — into the products as they ship. Duet is the first Anthropic product that looks like the thing he was describing: not a smarter chatbot, but an agent with its own computer, its own workflow, and a human checkpoint at every consequential step. The launch is less a product announcement than a live experiment in how much autonomy users actually want.
The playbook
- Design the agent around handoffs, not chats — Duet works on its own computer in the cloud and messages you when done or stuck — you hand work off and get interrupted only when a decision is needed.
- Gate the consequential actions — Messages, payments and document edits require your approval. The safety gate is the feature that makes autonomy usable.
- Win on connectors, not model alone — With 6,000+ connectors, the agent's value is access to your tools. The platform that wires into your stack wins the proactive-agent race.