{"slug": "what-claude-duet-needs-to-know-to-judge-what-s-urgent", "title": "What Claude Duet needs to know to judge what's urgent", "summary": "Anthropic opened a research preview of Claude Duet on October 21, a proactive agent built on Claude Opus 5.5 that sends a morning brief, checks connected apps every few hours, and works on tasks on its own virtual machine, available by waitlist to Pro and Max subscribers in the US. Anthropic says Duet is built \"to decide what needs you, what can wait, and when to check with you first,\" and it shows users exactly what it is about to do before sending a message, making a payment, signing a document or deleting anything. Anthropic acknowledges Duet \"might misread a thread or miscalibrate what's urgent,\" and the preview draws on more than 6,000 connectors in the Claude Marketplace plus the memory, skills and connected apps in a user's Claude account.", "body_md": "# What Claude Duet needs to know to judge what's urgent\n\nAnthropic opened a research preview of [Claude Duet](https://claude.com/resources/articles/claude-duet) on October 21. Anthropic calls it “a new app where Claude works as a proactive agent.” It comes to you with what needs your attention, keeps working when you step away, and checks with you before it acts.\n\n[Meta’s Muse](https://www.thisandthat.chat/blog/meta-muse-business-agent-learns-from-chats/) makes suggestions unprompted, and [OpenAI’s dots](https://www.thisandthat.chat/blog/openai-dots-always-on-agents/) bring you work “sometimes before you even think to ask.” Now Anthropic has built an assistant around the same idea, and we think all three are on the right track. An assistant that waits to be asked helps only when you think to ask, and much of the work in a company arrives in a message nobody planned for.\n\nWhat stands out in Anthropic’s announcement is where it puts the weight. “What makes Claude Duet so helpful is its judgment,” it says, and Anthropic built Duet “to decide what needs you, what can wait, and when to check with you first.”\n\n## What Anthropic built\n\nEach morning, Duet sends a short brief on the most important items across your connected apps. During the day it checks those apps every few hours and messages you only when something needs you. You can hand it tasks, from scheduling a meeting to building a sales forecast, and it works on them on its own virtual machine, so your laptop can be closed. It’s built on Claude Opus 5.5, can use the more than 6,000 connectors in the Claude Marketplace, and is open by waitlist to Pro and Max subscribers in the US. Anthropic built it for “professionals and small business owners with more on their plates than they can track.”\n\nDuet reads, drafts and prepares work on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it’s about to do and waits for your approval. That’s the right design. An assistant that prepares everything and acts on your approval gives you the chance to learn where to trust it.\n\n## Judgment needs context\n\nJess Provencher, a broker who owns Pro Homes, was one of Anthropic’s early testers. One day, without being asked, Duet flagged a property showing that was on her calendar for the wrong day. “[This is] exactly the kind of thing I want caught without having to ask,” she told Anthropic.\n\nTo make that catch, Duet needed two facts: the day on the calendar, and the day the showing was really for. It could compare them because both were in apps Jess had connected. Anthropic puts it plainly: “Your context from Claude is the foundation.” Duet begins with the memory, skills and connected apps in your Claude account.\n\nIf you work alone, or handle every conversation with a customer yourself, your own accounts hold most of what matters. At most companies, several people talk to each customer, so what the company knows is spread across everyone’s email and chat. Imagine a design agency with 30 employees. A client emails the account lead to say its launch has moved up a week, to November 12. The agency’s owner has the printer’s delivery date, November 14, in her own mail, and the old launch date on her calendar. When the printer writes that the delivery is on schedule, the owner’s Duet has no reason to flag it, because nothing in her apps says November 14 is now two days late.\n\nAnthropic is candid that Duet “might misread a thread or miscalibrate what’s urgent,” and says it will keep tuning Duet’s judgment based on what testers tell it. Some of those misses will be errors of judgment. At a company, many will come from facts that are only in someone else’s email.\n\n## Deciding what’s urgent takes facts\n\nDeciding what’s urgent is a question about what’s true now. The printer’s delivery is late only if the launch is on November 12, so an assistant has to know that the date changed, when it changed, and that the change still holds. A search of the mail could turn up the client’s email and an older one with the original date. The assistant still has to work out which one is current.\n\nThat’s what a company brain is for. It’s a set of pages about the customers, vendors, projects and decisions a company deals with. Each fact is tied to the message it came from and dated, so an assistant can tell what’s current.\n\n## A brain built from email\n\nWe wanted a company brain anyone could check, so we built one from Enron’s email. We gave Claude 45,047 messages from eight Enron mailboxes, and it wrote [214 pages](https://www.thisandthat.chat/enron/) about the companies, regulators and banks Enron dealt with. Each of its 49,401 facts quotes the email it was drawn from. Claude already knows a lot about Enron from what it read elsewhere, so a fact could come from that knowledge instead of the mail. To catch those, we discarded any fact whose quote we couldn’t find in its email word for word. When the mail shows a fact stopped being true, the page expires it on that date, and 19,528 facts expire. Claude wrote every page. No one edited a page or a fact.\n\nClaude Opus 5 and Opus 5.5 built and verified the pages, and Claude Fable 5.1 audited them. In a random sample of 500 facts, none was wrong at its core and 93% were fully supported by their email. The [methodology](https://www.thisandthat.chat/enron/methodology.html) covers the rest, including what the pages still get wrong and the hardest problem we hit: even when the quote was real, Claude sometimes added a detail from the public record that the mail never said, so a second pass had to check each fact against its email.\n\n## The assistant and the brain\n\nDuet decides what needs you. A brain gives it the facts it needs to decide. The two fit together: an assistant like Duet works from its owner’s apps and comes to its owner, and a brain reads the company’s messages and holds facts that everyone’s assistant can draw on.\n\nThat’s the layer we’re building at this+that. Our brain holds what a company knows, assembled from the messages it already gets across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, and kept current today by workflows that write back to it as they run, with a person approving or editing each update. We’re building toward a brain that keeps most of itself current from the message stream through one maintenance workflow, and the Enron pages are a prototype of it. DoBox turns the same messages into tasks for the person who should handle each one. Through our [MCP server](https://www.thisandthat.chat/mcp/), Claude can already read your brain and your DoBox.\n\nTry it on your own assistant. Find a date your team changed recently in a thread you weren’t on, like a launch or a renewal, and ask your assistant what’s urgent this week.", "url": "https://wpnews.pro/news/what-claude-duet-needs-to-know-to-judge-what-s-urgent", "canonical_source": "https://www.thisandthat.chat/blog/claude-duet-judgment-needs-context/", "published_at": "2026-10-29 10:18:14.428702+00:00", "updated_at": "2026-10-29 10:18:16.835924+00:00", "lang": "en", "topics": ["ai-agents", "ai-products", "artificial-intelligence", "generative-ai"], "entities": ["Anthropic", "Claude Duet", "Claude Opus 5.5", "Claude Marketplace", "Jess Provencher", "Pro Homes", "Meta", "OpenAI"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/what-claude-duet-needs-to-know-to-judge-what-s-urgent", "markdown": "https://wpnews.pro/news/what-claude-duet-needs-to-know-to-judge-what-s-urgent.md", "text": "https://wpnews.pro/news/what-claude-duet-needs-to-know-to-judge-what-s-urgent.txt", "jsonld": "https://wpnews.pro/news/what-claude-duet-needs-to-know-to-judge-what-s-urgent.jsonld"}}