Claude Duet: Give Proactive AI an Attention Budget

What changed with Claude Duet

Claude Duet is available through a waitlist, initially for Claude Pro and Max users in the United States. Anthropic says it works in its own virtual machine and browser, so monitoring can continue when a user's device is off. It can prepare a daily brief, watch connected accounts every few hours and surface items it considers important. [1][2]

The preview is deliberately constrained. Anthropic says Duet can read, draft and prepare work independently, but it waits for approval before sending messages, making payments, signing documents or deleting information. Some actions remain unavailable even with approval. The company also warns that the preview may misread threads or misjudge urgency, and that users cannot yet inspect its browser. [1]

Proactivity creates an attention cost

A reactive assistant consumes attention when a person asks for help. A proactive assistant decides when to claim attention. That changes the operating question from 'Did the task finish?' to 'Was this interruption worth the cognitive cost, and did the system have enough evidence to create it?'

A daily brief can be useful while still becoming noisy, repetitive or poorly timed. Monitoring can also widen the practical scope of connected data. Anthropic says Duet can work across connected apps and a catalogue of more than 6,000 connectors and plugins; that breadth is a capability claim, not a reason to connect everything at once. [1]

Set a four-part attention budget

Treat attention as a finite operating resource. Give each pilot a written budget with four controls:

• Observe narrowly: connect only the accounts and folders needed for one defined use case, with an owner and review date.
• Notify on thresholds: state which events justify an interruption, the maximum daily volume and which items belong in a scheduled brief instead.
• Prepare before acting: allow research, summaries and drafts, while keeping consequential external actions behind explicit approval.
• Record outcomes: log the evidence behind each alert, the user's response, any approval and whether the interruption changed a decision or prevented delay.

Run a controlled business pilot

Start with a low-consequence workflow such as meeting preparation, project risk summaries or follow-up drafting. For two weeks, measure useful alerts, ignored alerts, false urgency, repeated items, time saved and approval reversals. Review what Duet observed as well as what it produced. Remove connections that did not contribute to a useful outcome.

Do not treat silence as proof that monitoring is complete. Anthropic says Duet checks connected accounts every few hours because it shares plan usage limits, so time-sensitive coverage needs an explicit fallback. [1]

The practical opportunity is a calmer, better-prepared workday, not an assistant that interrupts without limit. ELYMENT AI helps organisations design approval boundaries, durable evidence and recovery paths so proactive automation remains accountable.

Sources

• Anthropic: Introducing Claude Duet (21 October 2026) - First-party announcement covering the research preview, availability, background monitoring, approvals, connectors and limitations.
• Anthropic: Claude release notes (21 October 2026) - First-party release record confirming the Claude Duet research preview and waitlist.

Continue learning

• Design a control boundary for agent monitoring
• Give AI agents runtime authority records
• Prepare an incident recovery plan for AI agents

Related analysis

• Claude Haiku 5.5: Test the Cost Threshold Before You Switch

  Claude Haiku 5.5 offers lower prices for short prompts. Test its 100,000-token threshold, quality, retries and escalation before changing business workflows.

Frequently asked questions