On October 21, 2026, Anthropic opened a waitlisted research preview of Claude Duet for US Pro and Max subscribers. Two days earlier, Notion shipped "Agents, rebuilt to do more for less". Together with OpenAI's dots, which we covered on September 29, the pattern is clear: the assistant is becoming a background worker with its own machine.
What Duet does
Anthropic describes Duet as "a new app where Claude works as a proactive agent." It is built on Claude Opus 5.5 and aimed at professionals and small business owners. It sends a morning brief, monitors connected apps, "has a virtual machine with its own browser, so it doesn't require your laptop or phone to be on", and messages you when work is done or needs a decision. The design details are the substance:
- Check-in cadence: it "checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster."
- Approval gates: before it "sends a message, makes a payment, signs a document, or deletes anything", it shows the action and waits. "A few things are off-limits even with your approval, like trading securities or moving money between accounts."
- Credential custody: it follows the Managed Agents design, so it "never directly sees your passwords or sign-in tokens". Purchases go through Stripe's Link with a virtual card.
- Leak control: a new classifier rejects requests to share "the strictest tier of information", such as card numbers and government IDs.
Anthropic is candid about limits: it "might misread a thread or miscalibrate what's urgent", and its background checks count against usage limits.
Notion makes the same move
Notion's October 19 release says: "We rebuilt our agent harness. Again." Its agents "now have their own computer and browser", a "Review for me" mode with safety checks, and "can work on tasks for up to 24 hours." Its Auto router claims results similar to Opus 5.5 and Astra "at about 90% lower average cost" across 20+ common tasks.
The reaction
Duet drew 322 points on Hacker News. One commenter said they have "a hard time making the trust leap". Another objected to model lock-in: "this feels like a product area where not being locked into one garden is more important". A third accused AI companies of "shipping their org chart" with a separate app.
What remains uncertain. There are no public measures of interruption quality, missed items or false alarms, and the 90% cost claim is Notion's own.
I think that once an agent has its own computer and runs for hours, the hard problems are an interrupt budget and credential custody. Duet's answers are worth copying in any internal agent: poll on a schedule you can afford, make irreversible actions explicit approvals, keep a short list of actions no approval unlocks, and keep tokens out of the model's context. The same week, LiteLLM's incident report described tokens leaking through logged response bodies, which is why that last rule matters. The product question to watch is calibration. An assistant that interrupts too often gets muted, and one that interrupts too rarely misses the call that mattered.
Claude Duet (research preview, October 21) runs Opus 5.5 on its own cloud VM, briefs users each morning, checks connected apps every few hours to conserve limits, requires approval before sending, paying, signing or deleting, and never sees passwords or tokens; Notion's October 19 rebuild gives its agents their own computer and browser for tasks of up to 24 hours, making interruption judgment and credential custody the design problems that matter.