A proactive agent needs an urgency ledger before an action ledger

Approval is not the only decision surface

Anthropic opened a research preview of Claude Duet on 21 October for waitlisted US Pro and Max subscribers. The product sends a morning brief, checks connected apps every few hours and can continue tasks in a cloud virtual machine. Before sending a message, making a payment, signing a document or deleting something, it shows the proposed action and waits for approval. Anthropic also says some actions, including securities trading and money transfers, remain unavailable.

The first-party post is unusually direct about an earlier decision: what deserves the user's attention. Duet is designed to decide what matters, what can wait and when to interrupt. Anthropic says the early version may misread a thread or miscalibrate urgency and is still being tuned. Background checks consume shared plan limits. The launch therefore changes two control surfaces at once: action authorisation and attention routing.

Reuters reported the preview, its connected-app monitoring and Anthropic's limitations. The Verge separately described the daily brief, approval points and dedicated sensitive-information controls. Neither report provides measured false-alarm, miss, completion or user-outcome rates. Early-user anecdotes in the vendor post show intended use, not general performance.

Independent interruption research supplies context, not a Duet test. A 2023 one-day field experiment with 247 participants found that disabling automatic communication notifications reduced interruption frequency and was associated with higher self-reported performance and lower irritation. Its short treatment, mixed worker-and-student sample and self-reported measures limit generalisation. It supports measuring interruption cost and role context; it does not establish how this preview performs.

Build an urgency ledger before expanding scope

A useful pilot should label a sampled stream of candidate items before letting the agent set the queue. For each item, record source, work context, urgency class, consequence of delay, consequence of a false interruption, accountable owner and latest useful review time. Ask a human to label the sample independently, then compare the agent's alert, defer and ignore decisions.

Report more than precision. Track high-consequence misses; low-value interruptions per user-day; time saved or added; duplicate alerts; stale summaries; sensitive-context exposure; and the share of alerts that arrived early enough to act. Segment by source and role. A sales lead, payroll exception, customer complaint and calendar conflict have different acceptable delays. One global “important” label erases that structure.

Test time as a variable, not merely a timestamp. The same item can be harmless at breakfast and critical ten minutes before a deadline. Include cases where later context reverses the initial ranking, where two alerts compete for the same owner and where silence is the correct outcome. Review whether the system explains why it interrupted and whether that explanation contains enough context for a fast human decision without leaking unrelated material. These tests reveal attention debt that task-completion metrics will miss.

Keep the action ledger separate. It should record the proposed action, data and tools involved, approval request, approver, edit, execution result and rollback path. A correct approval prompt does not repair a poor alert; a good alert does not justify autonomous execution. Join the two ledgers only through stable item and task IDs so reviewers can see whether the right signal reached the right person before a consequential action.

The counterargument is that manual labelling defeats proactivity. The ledger is a temporary calibration instrument, not permanent double work. Start with a bounded source set and a small group, review errors weekly and widen scope only when the miss and interruption profile is acceptable for that role. Preserve an off switch and a way for users to change what they hear about and how often.

This extends the earlier delegation register for always-on agents. That register answers what an agent may pursue; the urgency ledger tests what the agent decides to surface. For work-design leaders, the immediate decision is to treat prioritisation as a measurable capability with its own acceptance gate. An agent can ask before acting and still consume attention in the wrong places across the working day and across roles.