# Claude Duet wants to take the chasing out of everyday work

Agent contract: 1.2.0

A new research preview points toward AI that notices unfinished business and prepares the next step. The useful test is whether it gives people back attention while keeping them in control.

Canonical: https://brightaifuture.com/discoveries/claude-duet-proactive-ai-everyday-work
Format: discovery
Source publication: 2026-10-21
Bright publication: 2026-10-22
Substantive update: None recorded
Evidence and review: Experimental; confidence: unassessed; approved; ai-assisted. AI-assisted source review of current official launch, product FAQ, release notes, support, connector, agent-security and privacy guidance. Reuters corroborates announcement, not performance. Hypothetical workflows remain labeled; no Bright hands-on trial or independent benchmark.

## Editorial image

A laptop with a blank screen sits beside a phone, open notebook and pen on a sunlit wooden desk.: https://brightaifuture.com/media/content/52a5545cf37a8f354b2b9488c2de94057c5a8427482ef28965b096657d336b90.jpg

Credit: Contextual workstation photograph published October 3, 2025; not a Claude Duet screenshot, demonstration or verified workflow. Photo: Jakub Żerdzicki / Unsplash. Proportionally resized by Unsplash; no crop or other edit by Bright.. License: Unsplash License · https://unsplash.com/license.

Source: https://unsplash.com/photos/laptop-phone-and-notebook-on-a-wooden-desk-Qw_WS6EFY00

contextual; not Claude Duet, its interface, a person using it or a measured productivity result.

## The story

Most working days contain a second, less visible job: remembering what is still open. A client has not replied. A meeting needs preparation. An invoice needs checking. None of those tasks is necessarily difficult, but each occupies a small part of someone's attention until it is resolved.

That is the everyday problem behind Claude Duet, Anthropic's new proactive assistant. The interesting possibility is less time spent reopening apps to discover what needs doing, with more of the preparatory work already waiting when a person is ready to decide.

Anthropic opened the research preview on October 21, according to its [release notes](https://support.claude.com/en/articles/12138966-release-notes). Its [product FAQ](https://claude.com/product/claude-duet) says access begins through a waitlist for US Claude Pro and Max subscribers, with standalone iOS and Android apps and web access through claude.ai. Team and Enterprise plans are not included yet. [Reuters also reported the launch](https://au.marketscreener.com/news/anthropic-unveils-claude-duet-ai-agent-app-ce785cd9dd8ef527), describing the intended audience as professionals and small-business owners.

## What makes an assistant proactive?

A chatbot usually starts when someone asks it a question. A proactive assistant also takes responsibility for noticing when something might need attention, within the accounts and permissions its user has provided.

That adds two kinds of judgment. It must decide what work to pursue, and it must decide when interrupting a person is worth it. An assistant that flags everything merely creates another inbox. One that stays quiet about the wrong thing can leave a commitment uncovered.

The underlying agent pattern is a loop: plan a step, use a tool, inspect the result and adjust. Anthropic describes that distinction in its [framework for trustworthy agents](https://www.anthropic.com/research/trustworthy-agents). The model alone is only part of the system; its tools, operating environment and safeguards also shape what it can do.

In Duet's case, Anthropic says work runs on a cloud computer with its own browser, so a user's laptop or phone does not need to remain on. It can prepare work independently, then return when a result or decision is ready. These are [announced capabilities](https://claude.com/resources/articles/claude-duet), not a Bright hands-on finding.

## Where that could help in an ordinary week

Consider a freelance designer heading into a client meeting. A useful assistant might collect the latest brief, identify an unanswered question in the email thread and prepare a short meeting note. The designer would still judge the client's priorities, but would not have to reconstruct the conversation from scratch.

Or imagine a small consultancy tracking overdue invoices. The useful output would be a checked list of outstanding items, the relevant source records and draft reminders. A human would review whether each reminder is appropriate before anything goes to a customer, especially if a payment has recently arrived or a dispute is unresolved.

For a manager, the opportunity might be an unresolved scheduling thread. Rather than simply repeating that a meeting has not been booked, an assistant could collect the proposed times and prepare the next message. The manager's remaining job would be a clear decision, not another search through several apps.

A solo business owner might ask for a weekly sales summary. A genuinely helpful result would distinguish recorded figures from assumptions and make it easy to check the underlying records. Saving time on assembly is valuable; confidently producing an uncheckable number is not.

These are hypothetical workflows, not reports of Duet completing those exact tasks. They illustrate what to look for in a trial: less searching and chasing, useful preparation, and a short path from a claim back to its evidence. Success should be measured by work a person can confidently finish, rather than the volume of messages an agent produces.

## Permission to read is different from permission to act

Connecting an account is a meaningful decision. Anthropic's [connector documentation](https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities) says connectors can retrieve information and take actions, subject to the connected account's permissions. Third-party services process data on their own infrastructure and under their own terms.

That makes three questions worth separating: what information can the assistant access, what may it prepare independently, and what must a person approve before it changes anything outside the conversation?

For Duet, Anthropic says messages, payments, document signing and deletions require approval. Securities trading and transfers between financial accounts remain off-limits even with permission. Its [launch announcement](https://claude.com/resources/articles/claude-duet) also describes credential isolation, adversarial testing and safeguards intended to prevent inappropriate disclosure of sensitive information. Those are the company's descriptions of its protections, not proof that failures are impossible.

Approval should be an informed pause. Before sending a customer message, for example, a person needs to see the recipient and wording, not merely approve a vague instruction to handle it. Before paying, they need to check the merchant, item and amount. The point is to preserve the consequential decision while removing avoidable preparation.

Privacy needs its own check, too. Duet's [support guide](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet) says disconnecting an app stops future access but does not remove memories of information from earlier conversations. Anthropic separately provides [consumer model-improvement settings](https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings). Turning training off stops new chats and coding sessions being used for future model training, with stated safety-related exceptions. Access controls, memory and training preferences solve different problems.

There is also prompt injection: misleading instructions hidden in material an agent reads. Anthropic's [agent-security guidance](https://www.anthropic.com/research/trustworthy-agents) acknowledges that layered defenses cannot guarantee protection. A sensible starting point is a narrow, low-stakes workflow using only the accounts it actually needs.

## A preview with practical limits

Duet is not a continuous emergency monitor. Anthropic's [product FAQ](https://claude.com/product/claude-duet) warns that it can misread conversations, judge urgency incorrectly and miss something arriving between checks. That matters if a workflow depends on an immediate response.

The [support guide](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet) adds that users cannot currently watch or take over its browser, some Slack capabilities are limited, and Microsoft 365 is not available in Duet yet. The [announcement](https://claude.com/resources/articles/claude-duet) says personal errands, such as ticket booking, are weaker than its core work functions.

Cost also needs a practical reading. The [support guide](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet) says there is no additional Duet charge during the preview, but background activity counts toward the same five-hour and weekly limits as Claude chat. Included does not mean unlimited. Waitlist invitations arrive in waves, without a guaranteed access date.

## The progress worth watching

The hopeful part of this direction is quite human: fewer commitments held in memory, less repetitive coordination and more room to concentrate on the work that needs a person's judgment.

The evidence is still early. A launch establishes what a company is offering and the boundaries it describes; it does not establish how reliably the assistant will perform across different people's working lives. Useful independent testing would examine missed commitments, unnecessary interruptions, correction time and whether users can understand what happened.

For anyone admitted to the preview, a bounded trial is more informative than handing over an entire workday. Pick one recurring chore, check its outputs against the original sources and include review time when deciding whether it helped.

An assistant earns trust through a series of small, checkable successes. If proactive AI can deliver those while leaving consequential decisions with people, the reward will be more than a tidier inbox: it will be attention returned to the person who needed it.

## The human problem

Unresolved messages, meeting preparation and routine coordination occupy attention that people need for decisions and focused work.

## The prior constraint

People must reopen accounts, reconstruct context and track unfinished commitments themselves.

## AI’s actual role

A proactive cloud agent checks connected accounts at intervals and prepares tasks within user-granted access, with approval before consequential actions.

## The documented result

Anthropic opened a US Pro and Max research preview through a waitlist on October 21, 2026, describing background work, connected apps and approval gates. No independent reliability or time-saving result was verified by Bright.

## Why it may matter

Useful preparation could reduce searching and chasing while keeping consequential decisions with the user.

## Limitations

Research preview through a US Pro and Max waitlist; Team and Enterprise are not yet included.

Product capabilities and security safeguards are vendor descriptions, not independent performance validation. Bright has not tested Duet.

Duet can misread threads, misjudge urgency and miss items between checks; it is not continuous emergency monitoring.

Background work consumes shared five-hour and weekly limits. Included preview access does not mean unlimited use.

Users cannot currently view or take over its browser; Slack has limits and Microsoft 365 is not yet available.

Disconnecting an app stops future access but does not remove memories from earlier conversations. Training preferences, memory and account access are separate controls.

Everyday examples in this feature are hypothetical workflows, not demonstrations or testimonials.

## Unresolved questions

Does the assistant reduce missed commitments and coordination time after correction and review time are counted?

Can users understand and verify its choices, interruptions, source records and approval requests?

## Provenance and history

{
  "dates": {
    "captureDate": "2026-10-22",
    "eventDate": "2026-10-21",
    "lastReviewedDate": "2026-10-22",
    "publicationDate": "2026-10-21"
  },
  "provenance": {
    "origin": "editorial",
    "externalId": "https://claude.com/resources/articles/claude-duet"
  },
  "revisions": [
    {
      "id": "claude-duet-proactive-ai-everyday-work-initial",
      "recordedAt": "2026-10-22",
      "sourceIds": [
        "source:claude-duet-proactive-ai-everyday-work:1",
        "source:claude-duet-proactive-ai-everyday-work:2",
        "source:claude-duet-proactive-ai-everyday-work:3",
        "source:claude-duet-proactive-ai-everyday-work:4",
        "source:claude-duet-proactive-ai-everyday-work:5",
        "source:claude-duet-proactive-ai-everyday-work:6",
        "source:claude-duet-proactive-ai-everyday-work:7",
        "source:claude-duet-proactive-ai-everyday-work:8"
      ],
      "summary": "Initial source-checked feature on Claude Duet research preview, hypothetical everyday workflows, access limits and human approval boundaries."
    }
  ],
  "corrections": []
}

## Original sources

- [release notes](https://support.claude.com/en/articles/12138966-release-notes)
- [product FAQ](https://claude.com/product/claude-duet)
- [Reuters also reported the launch](https://au.marketscreener.com/news/anthropic-unveils-claude-duet-ai-agent-app-ce785cd9dd8ef527)
- [framework for trustworthy agents](https://www.anthropic.com/research/trustworthy-agents)
- [announced capabilities](https://claude.com/resources/articles/claude-duet)
- [connector documentation](https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities)
- [support guide](https://support.claude.com/en/articles/17631495-get-started-with-claude-duet)
- [consumer model-improvement settings](https://privacy.claude.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings)

## Continue exploring