Anthropic introduces Claude Duet, its artificial intelligence agent aimed at individuals. Still in ‘research preview,’ it is reserved for Pro, Max, and Ultra subscribers in the United States, with a waitlist. After a week of testing, journalist Allison Johnson describes it as sometimes very useful, sometimes a bit pushy.
An agent is a program that can perform tasks on behalf of the user, using the services connected to it. Duet is not for ordering meals: it is aimed, according to Anthropic, at ‘professionals and small business owners’ and must decide what most deserves the user’s attention. Like Meta’s Muse agent, it has its own app. Once messaging, productivity, and chat services are linked, it displays suggestions in the form of cards that can be dismissed or accepted with a swipe.
For the test, Allison Johnson gave it access to her work email, calendar, Google Docs, and Slack. Cross-referencing was quick. The team’s progress update, automatically summarized by email thanks to Gemini, could be linked to a private message on Slack. A shared document for weekly interviews also allowed it to identify her supervisor.
A useful assistant, but sometimes too zealous
Every morning, Duet sent a welcome message accompanied by a few key points for the day, like a reminder of equipment to bring for a shoot before a flight. It then suggested tasks, often followed by an offer to draft a message or add an item to an agenda. These suggestions were regularly useful, but the assistant sometimes gave the impression of being a slightly anxious secretary.
One Tuesday, the journalist had told an interviewee that the article would be published that day. She had submitted her text to her editors late Monday evening. Just before 9 a.m., Duet reminded her that the article was still awaiting a review. The alert was deemed unfounded: the agent did not take into account the clarification in the message stating that the deadline was not strict, nor the fact that it was only 7 a.m. for the editor.
At other times, the tool proved more relevant. While planning the following weeks, the journalist had not noticed that her upcoming days off would leave her with only a few working days before submitting a new major report. Duet proposed a solution, offered to complete the shared document with weekly notes for the editor, and then drafted a Slack message for them. The text imitated the journalist’s style, in lowercase, but it also shortened the word ‘signature’ to ‘sig,’ which she would never have done. The message also bore the note ‘sent with Claude.’
Duet also frequently suggested starting drafts, which, for the journalist, is an absolute refusal. The situation became more peculiar when the task was to review… this very test.
What the agent does not replace
Allison Johnson also noted the discomfort that can arise from having all her emails, drafts, and messages read. She accepted this data sharing knowingly, but seeing the tool in action is different. She wonders about the effect such an assistant would have on her way of working, questioning whether she would end up feeling detached from her professional environment.
For her, the tool mainly clarified what she expects from AI. Spotting forgotten inconsistencies between emails, messages, and documents is useful, and the idea of a safety net for her organization appeals to her. On the other hand, she rejects AI in the creative phase. She cites Hank Green, who describes his own use of AI as a way to find ‘my own paths in and around a subject.’ According to her, entrusting groundwork to a robot leads to losing unexpected discoveries.
