Claude's new agent checks your accounts every few hours. How often it looks is the next thing you'll pay for.
26 Oct 2026 · News
The personal agent race has a new test. Everyone agrees the agent can act. The hard part is knowing when to speak up, and that turns out to cost money.
Anthropic's new personal agent, Claude Duet, looks at your email, calendar and other connected apps every few hours. Anthropic says why in one plain sentence: checking nonstop "would use up your plan's limits much faster." So how quickly your agent notices something is a cost, and costs get priced. Our call: by 30 June 2027, at least one of OpenAI, Anthropic, Google or Meta sells more frequent background checks as part of a higher paid plan.
What Anthropic actually said
Duet opened as a waitlisted research preview for US Pro and Max subscribers. Its launch page pitches judgment first. Each morning it sends a short brief, and through the day it "messages you only when something needs your attention."
Further down, the page is unusually candid about how that works. Duet "checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster." Those background checks count against the same usage limits as the rest of Claude. Anthropic adds that it is "continuing to test its judgment about when to message you," and that the preview "still might misread a thread or miscalibrate what's urgent." You can tell it what you want to hear about and how often.
In other words, the agent can't watch for you all the time, because watching uses the same model time as everything else you do with Claude.
Two ways to get it wrong
An agent that interrupts you can fail in two directions. It stays quiet when something matters, or it keeps raising things that don't.
The first hands-on reports show both. On the 25 October episode of The AI Daily Brief, the host described his OpenAI dot as "not particularly proactive." After he shared roughly six hours of strategy conversation with it, the reply amounted to "great, this information has been saved." Duet went the other way. It decided two sponsors booked on the same date was a conflict, which it wasn't, and reminded him "like four times" until he told it to drop the whole topic. He still preferred Duet, because its follow-up drafts were good enough to send.
That's one person's experience. It matches what Anthropic's own page warns about, so we'll take it as the shape of the problem.
Then versus now
On 30 September we asked what a dot does when you aren't asking it anything. OpenAI's answer was "proactive research," read-only background work on the apps you connect. Four weeks in, the early answer from users is "not much."
On 21 October we argued that you should judge an agent on whether it pushes back, not on how warm it sounds. Restraint is the second test, meaning whether it knows what to leave alone. Restraint is also where the bill shows up, because every look at your inbox to decide whether to speak up uses model time.
And on 12 October we called the flat AI subscription a meter you can't see. Duet adds a new thing that meter counts. Your plan now pays for the agent's attention as well as your chats.
Why it gets priced
There's a fair case that this sorts itself out for free. Labs copy each other's features within weeks, and a passive dot can be tuned to speak up more. But tuning when the agent talks is cheap, and the looking that comes before it costs model time. Checking every hour instead of every four takes about four times as many checks, whether or not anything is waiting.
OpenAI has already pointed in this direction. Its dots announcement says that "in the future" you'll be able to "scale the output of each dot by either increasing its speed or the total amount of work it can take on per month." So OpenAI already plans to charge for speed and monthly workload. How often the agent checks is the obvious next upgrade, and it's an easy one to sell, because waiting four hours to hear about a moved meeting is something people notice.
Our call
By 30 June 2027, at least one of OpenAI, Anthropic, Google or Meta will name background check frequency, or how fast the agent reacts to new mail and calendar changes, as a difference between its plans. Faster checking costs more. We'll count ourselves wrong if mid 2027 arrives with real-time watching included in every paid plan, or with check frequency still unpublished and the same for everyone.
Until then, the most useful setting on any of these agents is the one Anthropic mentions almost in passing. Tell it what you don't want to hear about. Each topic you rule out is one less false alarm.