Anthropic unveiled Claude Duet on Wednesday, a standalone agent app that does not wait to be asked. Announced on October 21 and opening as a waitlisted research preview for United States subscribers on the Pro and Max plans, Duet monitors the apps you have connected, sends a briefing each morning, checks back every few hours, and alerts you when something needs your attention. It is built around Anthropic's Opus 5.5 model, runs on iOS, Android and claude.ai, and works in one ongoing thread that reuses the memory, skills and connectors from your Claude account.
The work Duet is pitched to do is deliberately ordinary, and that is the point: scheduling meetings, chasing unpaid invoices, drafting follow-ups, and building financial forecasts, dashboards and even web apps. It runs on its own cloud computer with its own browser, so it keeps working after you close the app. Anthropic says it draws on more than 6,000 connectors in the Claude Marketplace, including Gmail and Slack, and can make purchases through Stripe's Link wallet and Shopify's Shop Pay. The guardrails matter as much as the capabilities: Duet is designed to ask before it sends a message, makes a payment, signs a document or deletes a file, and it does not see your passwords. Anthropic is candid that it can misjudge urgency, is weaker at tasks such as booking tickets, and its background checks count against your plan limits.
Duet is not an idea in isolation. OpenAI launched its always-on Dots agents last month to pursue goals across apps such as Slack and Microsoft Teams, and Meta's Muse agent has already drawn millions of downloads. All three labs have converged on the same bet: the next interface is not a chat box you visit, but a delegate that watches, prepares and pings you for approval. For builders, that shifts the hard problems from prompt quality to permission design, audit trails and cost control. Duet is not yet on Team or Enterprise plans and Anthropic announced no Duet API, so for now it is not a platform to build on directly.
The same day offered the other half of the agent story. TechCrunch reported that OpenAI and Google's Wiz used AI to find two flaws in PHP that had sat in every version since PHP 5 in 2004, in a language that powers roughly 70 percent of websites. Calif, working in OpenAI's Patch the Planet initiative with GPT-5.6 Sol, found one flaw; Wiz, using its Atlas agent alongside Gemini 4 Argon and GPT-5.6 Sol, found more. The PHP Foundation has issued patches, the current release is 8.5.12, and versions 8.1 and older are unsupported. The NSA warned on Wednesday that a site that cannot be patched should be taken offline until it can. OpenAI says there were no reports of exploitation in the wild, but Wiz's warning stands: flaws once protected by being too hard to exploit lose that protection when models can adapt an exploit across versions and configurations.
Put together, the lesson for anyone running websites or building with AI is practical. Patch PHP now: if you run Drupal, WordPress or any PHP stack, confirm your server version and apply the Foundation's fixes through your host before disclosure becomes an attack script. And if you try a proactive agent like Duet, start it read-mostly, with briefings and drafts first and purchases and sends only after you trust its judgment about what is urgent. The tools are getting better at acting without being asked. Your job is to decide, in advance, exactly where that autonomy stops.
