Anthropic Launches Claude Duet Agent App for Professionals

Full analysis

The Skeptic. Six thousand connectors is a brochure, not a capability. The long tail is stale OAuth, undocumented endpoints, and plugins that haven't been touched in a year. Day one, a meaningful chunk returns garbage the model has to guess at. The "professional, won't book your tickets" framing is positioning, not engineering. Either Opus 5.5 handles multi-step tool chains reliably or it doesn't, and the task category doesn't change that answer. There's no public benchmark on tool-use accuracy under messy conditions, and no production users. Three agents launched in one week by three labs with no traction between them. That's a race to plant a flag, not proof anybody solved the hard part.

The Safety Lens. Blocking credit card and government ID requests is a real design choice, but it guards a narrow door. The open door is prompt injection through connectors. A poisoned document in your inbox or a malicious plugin can steer the agent into buying or sending things, and no PII filter catches that because no PII crosses the line. Anthropic's ban on securities trading and fund transfers is sensible, but it's a list of forbidden actions, not a boundary the model understands. Attaching Shop Pay and Stripe Link to an agent that reads untrusted documents is the exact setup injection researchers have been warning about. The waitlist is the right posture. The pressure from Meta and OpenAI will shorten it faster than the adversarial testing can keep up.

The Compute Pragmatist. A background agent on Opus 5.5 is a different cost animal than chat. Long context held open, retry loops, multi-step tool calls, task queues that don't batch cleanly. Parsing 6,000 kinds of API response burns tokens on reasoning before any work gets done. Cost per active Duet session will run well above cost per chat, and Pro/Max subscription pricing was never built for a user who leaves an agent running all day. If adoption lands, margins bleed quickly. That's why this is capped behind a waitlist and called a "research preview." It's not just safety caution. It's Anthropic not knowing what a heavy user costs them yet, and not wanting to find out at scale with a flat monthly price.

The Enterprise Buyer. The thing a small business would actually pay for, an agent that buys supplies and builds the Monday briefing, is also the thing procurement can't sign. Who's liable when the agent buys the wrong thing on Shop Pay? Where's the audit log of every autonomous action? What happens to data flowing through a third-party connector nobody vetted? "Surfaces key decisions for approval" is a liability dial, and the preview doesn't tell me how it's tuned. For a solo professional this is a convenience. For anything with a compliance function, the autonomous-purchase feature is a reason to wait, not a reason to buy.

The Builder. The approval UX is the whole product, and nobody has solved it. Interrupt too often and the user abandons the agent as more work than doing it himself. Interrupt too little and it spends money it shouldn't. That dial is nearly impossible to set right. On the plumbing: whether Marketplace connectors return standard error schemas or freeform failures decides everything, because a freeform failure means the model has to interpret what broke, and that's where background tasks silently die. Anybody building on Claude's APIs should assume the 6,000-connector surface leaks error states the demo never showed.

Where they disagree. The Compute Pragmatist and the Skeptic are looking at the same fact from opposite ends. The Skeptic says the waitlist means no product. The Pragmatist says the waitlist means Anthropic can't yet afford the product at subscription pricing. Both read the gate as weakness, but one is about capability and the other about unit economics, and they point to different fixes. The Safety Lens and the Enterprise Buyer agree the money-attached actions are the problem, and disagree on nothing except urgency. The real split is between the Builder, who thinks the approval dial is solvable with enough iteration, and everyone else, who thinks the dial IS the unsolved problem and iteration doesn't close it fast.

What it hinges on. One question: does an always-on agent with real purchasing power survive contact with untrusted inputs and messy connectors without either annoying the user into quitting or spending money it shouldn't. Nobody has shown that yet, from any lab. The waitlist across all three launches is the evidence that nobody can. Three frontier labs shipped the same unfinished product in one week because the category is a land-grab, and the gating tells you the reliability and cost work isn't done.

Prediction: Claude Duet will still be gated behind a waitlist or labeled "research preview" at Anthropic's next major model release (Opus 6 or equivalent), rather than generally available to all Pro and Max subscribers.

Confidence: Medium. The money-attached, injection-exposed design makes fast general availability reckless, and the pricing math discourages it.

Why: Duet attaches real purchasing power (Stripe Link, Shop Pay) to an agent that reads untrusted documents and calls 6,000 third-party connectors, which is the exact setup where prompt injection turns a poisoned inbox into an unauthorized purchase, and no PII filter catches that. Opening this to every subscriber before that's hardened invites a public incident Anthropic's whole brand is built on avoiding. On top of that, a background agent on Opus 5.5 costs materially more per session than chat, and flat Pro/Max pricing wasn't built for it, so uncapped general availability exposes margin before the economics are known. The opposite (fast full general availability) only happens if Meta Muse or OpenAI Dots forces Anthropic's hand, and even then Anthropic has historically chosen slower, gated rollouts over racing to match a competitor.

Revisit by 2027-06-30: We're right if Duet is still waitlisted or preview-only, or still restricted to a subset of subscribers, at Anthropic's next flagship model launch. We're wrong if Duet is generally available to all Pro and Max subscribers with autonomous purchasing on, no waitlist, by that launch.