2026-10-22 18:26 UTC
DANGMUAAI & Developer Tools, Decoded
BackAgents

Claude Duet: Anthropic's Always-On Agent Hits Preview

Anthropic's Claude Duet works between your prompts on Claude Opus 5.5, with 6,000+ connectors, a cloud VM, and an approval gate on every sensitive action.

DangMua EditorialOct 22, 20265 min read
Claude Duet: Anthropic's Always-On Agent Hits Preview

Anthropic launched Claude Duet on October 21, a research-preview agent that works between your prompts instead of waiting for them. It runs on Claude Opus 5.5 and is waitlisted to US Pro and Max subscribers.

Duet joins a cluster of always-on agents shipped in roughly six weeks, and it draws the tightest line of any of them around what an agent may do unsupervised.

What Duet actually does

Duet reads your email and calendar overnight and sends a morning brief on what needs you, with replies already drafted. It then re-checks the apps you connect every few hours. Between those checks it works on its own cloud virtual machine with a browser, so tasks continue while your devices are off — building dashboards, forecasts, invoice chasing, payroll prep and decks, per the launch write-ups.

Connectivity comes from the Claude Marketplace: over 6,000 connectors and plug-ins, including Gmail, Slack, QuickBooks, Stripe, Gusto, DocuSign, HubSpot and Shopify. Duet can also make online purchases on your behalf through Stripe's Link digital wallet or Shopify's Shop Pay.

Anthropic's own launch examples are deliberately unglamorous and aimed at small businesses: a $640 county permit fee held for approval, a contract redlined to Net 60, and 27 wrong delivery orders with refunds already drafted.

The approval gate is the product

Duet will not send a message, make a payment, sign a document or delete a file without your approval. Anthropic says it will not trade securities or move money between accounts at all, and that Duet never directly sees your passwords.

The company also shipped a classifier that inspects each message to stop sensitive information leaking out. By default the agent declines requests for credit card numbers and government identification.

Read against the category, that is a positioning choice, not a safety footnote. Anthropic explicitly distinguishes Duet from personal agents like Instinct by saying Duet is not good at errands or booking tickets, and that it cannot show you its browser use as it happens. The pitch is a bounded worker for professionals and small businesses, not an assistant that does everything and shows its work.

The limits Anthropic printed itself

Three caveats come straight from the launch, and they matter more than the demo list:

  • Duet "can still misjudge what's urgent" — Anthropic's own wording for a product whose core output is a prioritised brief.
  • Background checks consume your existing plan usage limits. An agent that polls every few hours is spending the same quota you use interactively.
  • Access is narrow: US Claude Pro and Max subscribers, waitlist only. Team and Enterprise plans are not included yet, with no dates given.

The second one deserves attention from anyone sizing this for a team. The cost of an always-on agent is not a new line item — it is drawn from the quota you already budgeted for hands-on work, and polling frequency is what sets the burn.

Where it sits in a crowded six weeks

Duet lands into a field that filled up fast. One launch write-up counts it as the fourth always-on agent in about six weeks. Meta's Muse arrived in September with the widest consumer reach; OpenAI shipped Dots at DevDay on September 29, with a reported 4,000-plus apps; Google has been rolling agent features through its products. TechCrunch also lists Hark, Instinct, Grok Bot and Fo among the rivals Duet is aimed at.

The split in the category is already visible. Meta's Muse is a personal assistant, though Meta also released a version for small businesses with connections to other apps. OpenAI's Dots carries a business slant through its integration with professional tools. Anthropic has gone furthest toward work and furthest from errands.

By reach, Duet is the narrowest of these launches — one country, two paid tiers, a waitlist — while running the flagship model underneath and applying the strictest approval gate. One of the launch write-ups makes the same read, calling it the smallest launch with the flagship model and the strictest gate. That combination is a bet that the blocker on agent adoption is trust rather than capability.

What to watch

Duet is available as an iOS and Android app and on the web through Claude.ai. Three things will tell you whether the bet is working before any usage numbers do.

First, whether the approval gate survives contact with volume. A gate that fires on every payment is a feature at ten tasks a week and a nuisance at a hundred; the design that wins will be the one that learns which approvals you always grant.

Second, whether usage-limit consumption becomes the real pricing conversation. Background polling turns a subscription into something closer to metered compute, and the plan that fits a five-person business is not obviously the plan that fits one owner.

Third, whether Team and Enterprise access arrives with dates attached. Anthropic aimed Duet at professionals and small businesses but shipped it on consumer plans only — the gap between those two facts is the clearest unfinished edge of this launch.

More from DangMua