2026-11-13 · ← News
Claude Duet can watch your work, but it still confuses urgency with morning anxiety
A week with Claude Duet showed useful oversight across email, calendar, documents and Slack, alongside weak judgment about tone and real urgency. The proactive agent faces a harder problem than completing tasks: knowing when to leave its user alone.
The image could not be loaded.
The Verge tested Claude Duet for roughly 7 days in a normal work routine. Anthropic’s agent connected email, calendar, Google Docs and Slack, produced a morning brief and offered suggested tasks as cards that could be accepted or dismissed.
Duet assembled scattered work signals into one morning brief
In the test, it connected a meeting detail summarized in an email with a direct message in Slack. It inferred who the writer’s manager was from a shared document, spotted a narrow work window before a deadline, proposed a revised plan, updated the document and drafted a message.
That is where a proactive agent can be useful. Modern work is split across services, and people often miss a relationship hidden in another thread rather than the task itself.
The real product is attention judgment, not another text generator
Duet does more than answer a request. It watches connected services, decides what deserves attention and proposes the next action. Anthropic built it around Claude Opus 5.5, with confirmation before it acts on a user’s behalf.
For knowledge workers, this changes the interface. Instead of searching individual applications, they receive an evolving list of exceptions. The value depends less on whether the agent can draft a message and more on whether it recognizes the right exception.
The agent read the context but missed a soft deadline and the time zone
Before 9AM, Duet pressed the writer about an article awaiting edits, even though it was only 7AM for the editor and the promised date was not a firm deadline. Elsewhere it copied the writer’s lowercase style but invented an abbreviation she would never use. Pattern recognition is not social judgment.
The convenience also requires broad access. In this test the agent could read work email, documents and Slack. The writer opted in, yet still found the practical experience of continuous observation uncanny. Permission management and an auditable action trail will matter as much as model quality inside a company.
Early access is narrow, and bad interruptions will decide the outcome
Duet remains a research preview for waitlisted US Pro, Max and Ultra subscribers. It runs in standalone iOS and Android apps and on claude.ai. Background checks consume the same plan limits as other Claude use, so proactive monitoring competes with regular conversations and tasks.
The useful metric will not be how many tasks Duet proposes. It will be the share of alerts accepted without correction and the frequency with which it raises an alarm over a soft deadline or a misunderstood relationship.
Lilith's verdict
A personal agent gets a seat in every work conversation. If it wakes the editor at 7AM over a soft deadline, it has not delivered judgment. It has automated anxiety.