Anthropic has announced Claude Duet as a research-preview agent for professional work, according to The Verge. Its report describes a US waitlist for Pro or Max subscribers, connected-app briefings, background tasks and approval requests before consequential actions such as payments or deletion. It also relays Anthropic’s warning that this early version can misread a thread or misjudge urgency.
Engadget reports that Duet shares usage with other Claude tools and checks connected accounts every few hours rather than continuously, to conserve plan limits. It says Anthropic plans further safety information. These are reported product descriptions; they do not establish independently measured reliability.
The interval between checks is especially important. A system can initiate a briefing without observing every event as it happens. Calling that system proactive describes who starts the interaction; it does not establish how quickly a new event will be noticed. Those are separate properties of the service.
A check interval is part of the workflow
For a hypothetical office task, imagine a document changing just after an account check. A later review might identify the change correctly, but the time before that review still matters if someone needed the information earlier. This is an illustration of periodic checking, not a measurement of Duet’s actual latency or a promise about its schedule.
A useful evaluation would therefore record the time an event becomes available, the time the agent sees it and the time a person receives a message. Measuring only whether the eventual summary is correct would miss the first two intervals. An accurate alert can still be late for the particular task it is meant to support.
The reported phrase “every few hours” is not an exact service-level commitment. It should not be converted into a guaranteed maximum delay. Nor does it tell a reader whether all connected services are checked together. Those implementation details remain questions for documentation or an observed test.
Notification judgment is another decision
After noticing an event, an agent must decide whether to interrupt someone. This creates a different evaluation problem from reading the event accurately. A team might agree with the summary while disagreeing with its priority. If the assessment records only successful messages, it will omit events the agent saw but chose not to surface.
A proposed acceptance exercise could use a small set of clearly described work items with expected priorities agreed in advance. Record both alerts and non-alerts, then compare them with that reference. The purpose would be to examine notification judgment under stated conditions, not to create an all-purpose intelligence score.
Anthropic’s reported warning about urgency makes that distinction relevant to the preview. It does not provide an error rate, and this article does not invent one. A quiet day could mean nothing important happened, or it could mean something was not surfaced. Without an event record, silence alone cannot distinguish those explanations.
Approval does not resolve every earlier choice
Requiring a person to approve a consequential action creates an opportunity to review it. That opportunity is valuable only to the extent that the person can understand what is proposed. A useful assessment should inspect the proposed action and supporting context rather than treat the existence of an approval button as the complete result.
In a hypothetical scheduling task, a meeting invitation could be correctly drafted after the agent has misunderstood the purpose of the conversation. A final approval request might allow the person to catch that problem, but it would not mean the earlier interpretation was correct. Interpretation, proposal and authorization should be assessed separately.
This is not a claim about a defect observed in Duet. It is a reason to evaluate a connected-work assistant across the whole chain of decisions. A product can ask permission at the end while still needing careful review of what led to the request.
Shared usage changes the comparison
Because Engadget reports a shared usage allowance, a buyer should avoid treating background work as a separate unlimited service. The current reporting does not provide enough information to calculate a particular user’s remaining capacity or an effective price per completed task. That calculation would need actual usage rules and the person’s other Claude activity.
A practical preview assessment could record the work requested, the useful outcomes and the allowance consumed during the same period. Keep those observations separate from a prediction about general availability. The announcement’s access conditions and the eventual service terms are not automatically interchangeable.
Duet’s reported design therefore raises a concrete operational question: which tasks tolerate periodic observation and human review, and which require a different response schedule? The announcement identifies a direction for the product, while leaving its performance under a particular workload to be demonstrated.
Before relying on a proactive agent for time-sensitive work, define the latest useful alert time and check whether its documented observation schedule and approval flow can support that requirement.
