Anthropic 披露 Claude Duet 的优先级评分与独立意图校验机制

Anthropic built Claude Duet, an AI assistant that monitors your connected apps and works on tasks in the background, to decide what needs your attention without being asked. A week after its October 21 debut, the frontier lab with an IPO in its sights has shared with The New Stack how it is testing whether Duet gets those priorities right.

Anthropic has developed internal graders to assess how Duet prioritizes tasks, though urgency can be subjective. The stakes rise when an agent manages work across email, calendars, and B2B apps. Duet can remain within its permissions and still get its priorities wrong.

Powered by Claude Opus 5.5, Duet can prepare work and flag items it thinks need attention without waiting for instructions. Anthropic has been testing that approach with its own employees, using their day-to-day work to see how Duet handles competing priorities.

Testing Duet’s judgment

“After using Duet for our day-to-day work at Anthropic, we calibrated graders that assess correctness in task prioritization, and we run them offline against cases representative of internal use and early external testing,” a company representative tells The New Stack.

Those graders measure whether Duet can distinguish what deserves attention from what can wait, though Anthropic acknowledges the difficulty of evaluating those decisions.

“Urgency is an inherently subjective call, so we found building an eval for it imperfect but quite helpful,” a company representative says.

Duet can spot tasks like an approaching deadline or an email requiring a reply, but it doesn’t always know which task should be a priority, and the company admits the agent can misread conversations or misjudge urgency. Anthropic’s public launch and support materials do not say how often that happens.

Testing Duet’s safeguards

On October 23, two days after Duet’s debut, Anthropic published a closer look at its safety, privacy, and security safeguards. Duet team lead Karan Ahuja pointed to the post while addressing feedback received through messages, emails, and public posts.

We’ve been humbled by the interest in Claude Duet, and are working hard to get people off the waitlist. In the meantime, a few thoughts based on DMs, emails, posts we’ve seen.

→ Where: You can use it on iOS,… pic.twitter.com/9MMMZbHpLw— karan ahuja (@ahuja_karan) October 23, 2026

Anthropic keeps the part of Duet that does the work separate from the part that checks whether the work matches what the user actually asked for.

“A separate, automated check runs alongside each action, built on the auto mode classifier from Claude Code and extended for Claude Duet,” the rep tells The New Stack. “It assesses whether the action matches your intent.”

Because that check runs independently of the model performing the task, Anthropic says instructions hidden in a webpage can’t directly steer the classifier’s decision.

Anthropic says an outside firm penetration-tested Duet, but its public materials don’t show how reliably the separate check catches actions outside the user’s intent. Related research illustrates the difficulty of enforcing this agent scope: Researchers found that removing a prompt section defining authorized and out-of-scope actions in Claude Code tests increased out-of-scope actions, although that study did not test Duet’s separate classifier.

Limiting prompt injection

The separation is designed to reduce prompt injection, where malicious instructions embedded in emails, documents, or webpages try to redirect an agent. Researchers have already explored this risk by leaving hidden messages for rogue agents and observing how they responded. Anthropic’s support materials still tell users to review drafts and approval requests because Duet may encounter prompt injection.

Duet uses the apps already connected to a user’s Claude account. Disconnecting an app stops new access, but information Duet has already stored in memory remains until the user resets it. Anthropic says Duet never sees passwords, browser cookies, or sign-in tokens, and can’t save them to memory.

Duet is designed to seek approval before actions such as making payments or deleting items. Sharing private health or financial information requires consent, though an explicit request to share it can count as that consent, while trading securities, moving money between bank accounts, and accessing someone else’s account remain off-limits entirely.

Summaries, not surveillance

Users can see what Duet has finished and what’s still underway through its task list, with a separate page for each job. They can also give the agent new instructions while it’s working.

But users can’t see everything happening behind the scenes. During the research preview, users can’t watch Duet work in its cloud browser or take over the browser. Users can review its task activity and approve actions, but they can’t inspect each browser step as it happens. (The public launch and support materials do not describe a way to undo completed actions across connected apps.)

What background work costs

“Larger tasks, like building a financial model or a deck, will require more model turns and background work,” the Anthropic rep tells The New Stack.

Even Duet’s routine checks across connected apps count toward the five-hour session and weekly limits.

The company expects to keep adjusting the product as testing continues.

The post “Urgency is an inherently subjective call”: Anthropic tells TNS how it grades Claude Duet appeared first on The New Stack.