Claude can keep working while your laptop is off
Duet’s US preview has a waitlist—and background tasks draw from the same limits as your regular chats.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:004:56
Read transcript
Claude can now keep working after you close the chat—and even when your laptop is off. Anthropic has opened a waitlist for Claude Duet, a research-preview assistant that runs on its own cloud computer and browser. It can check connected accounts, prepare work and continue tasks while you’re away. The aim is to move Claude from answering prompts to handling ongoing work between them. Duet is included for U.S. Pro and Max subscribers, through standalone iOS and Android apps and claude.ai. Invitations are rolling out in waves, with no guaranteed access date; Team and Enterprise subscribers aren’t eligible yet. Duet inherits your Claude account’s memory, skills and connected apps. It sends a morning brief, then checks those accounts every few hours. Larger assignments can come back as editable research briefs, financial models, slide decks or dashboards. A separate Tasks tab shows work that needs your input, and only you can mark a task complete. The important boundary is that reading, drafting and preparation can happen without another prompt, but Anthropic says Duet must wait for approval before sending messages, making payments, signing documents or deleting information. Securities trading and transfers between accounts are prohibited even with permission. Anthropic also warns that emails and webpages can contain malicious instructions, and says no defense against them is perfect. During the preview, users can’t watch or take over Duet’s browser, and the company acknowledges it may misread a thread or misjudge urgency. There’s no added fee, but there’s no separate usage allowance either. Background checks and briefs draw from the same five-hour and weekly limits as ordinary Claude chats, so Duet can use up your allowance sooner. And checking every few hours means urgent messages may sit until the next check. The practical test is whether the time saved outweighs those limits—and whether you’re comfortable delegating preparation without continuous visibility. That shift from answering to doing has a parallel in research: Microsoft says AI agents helped researchers prove the KLS conjecture, a 31-year-old problem in geometry and probability. Xinzhi Zhang and Zhao Song directed the work, exploring more than 100 possible approaches with AI agents. They selected the promising directions, assembled the argument and verified the proof; the result appeared October 4. The conjecture concerns bottlenecks in high-dimensional probability distributions, which affect how quickly random walks can explore them. A later Lean 4 formalization combined the researchers’ criterion with additional research and machine-checked an argument. The distinction matters: that check supports the mathematics, but the account of how guided AI compared with less-directed attempts comes from the researchers themselves. The payoff is that results previously stated with KLS as an assumption can now be stated without it. And as AI moves into work, payroll data offers an early signal of who may be feeling the change. Stanford and ADP’s Canaries Dashboard found that U.S. employment among 22- to 25-year-olds in the most AI-exposed roles was down 4.4 percent year over year in August. Those roles have contracted monthly since October 2023. That’s a warning signal, not proof that AI caused the decline. The dashboard uses anonymized payroll records to track selected occupations, rather than representing the whole labor market. It also shows slower job growth for young women than young men, but Stanford says the data may not be representative enough to confirm that pattern. Longer-term government statistics show no evidence of a gender gap. The useful watch is whether this narrow early signal holds up in broader data. The same gap between AI’s promise and the need for safeguards shows up in a PHP security update. PHP has patched four critical flaws after Wiz researchers demonstrated server takeover without logging in on WordPress, Drupal and Joomla deployments. The bugs are in PHP itself, not necessarily in a site’s own code; Wiz says even a simple application on an affected PHP version could be vulnerable. The flaws can also expose sensitive process memory, including passwords, tokens or API keys. Wiz says AI helped researchers adapt working exploits across applications in hours, by analyzing failures and revising attacks through testing. The company calls the PHP update the only complete remediation. PHP’s broad use—estimated by the Wall Street Journal at as much as 70 percent of websites—is not a count of vulnerable sites: exposure depends on versions and deployment details. For defenders, the concrete next step is to patch, then investigate unusual multipart requests, repeated server errors, malformed PHP session cookies or unexplained worker-process churn. Across these stories, AI is increasing the pace of both useful work and consequential risk; the thing to watch is whether human review, reliable evidence and operational controls keep pace.
