AgentsAutonomy & tool use
Anthropic launches Claude Duet, a proactive Opus 5.5 agent for US Pro and Max subscribers
Anthropic's Claude Duet runs on Opus 5.5 in a private cloud computer, checks connected apps every few hours, sends morning briefs and asks approval before acting. It follows rival agents from xAI, Meta and OpenAI and opens to US Pro and Max subscribers via waitlist.
From the TechCrunch AI article of October 22, 2026 and cited sources: Duet works in the background on a virtual machine with its own browser, and it keeps running when the user's devices are off. [1] [9] [11] [12]
SOURCES[1] TechCrunch: Claude Duet app is Anthropic's bet on agents techcrunch.com[9] MacRumors: Anthropic Launches Claude Duet macrumors.com[11] TestingCatalog: Claude Duet testingcatalog.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Duet works in the background on a virtual machine with its own browser, and it keeps running when the user's devices are off. [1] [9] [11] [12]
- Original source
- TechCrunch AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 22, 2026 article published
- Covered
- Oct 23 edition
- Evidence
- 12 sources across 11 sites
- 1TechCrunch: Claude Duet app is Anthropic's bet on agentstechcrunch.com
- 2Claude Duet App: 4 AI Giants Built the Same Agent in 71 Daysreddit.com
- 3Claude Duet: Claude’s New AI Agent Is Herereddit.com
- 4AI Weekly: Claude Duet waitlistaiweekly.co
- 5Linked video: Claude Duet app is Anthropic’s bet on agentsyoutube.com
- 6TechNews: Opus 5.5 launchtechnews.tw
RisingHow it’s measured
YouTube video · 864.8K views · 10.1K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
Duet works in the background on a virtual machine with its own browser, and it keeps running when the user's devices are off. It connects to more than 6,000 apps and reuses the memory and skills already in the user's Claude account. It checks apps every few hours rather than in real time. It is built for briefings, dashboards and suggested actions. It handles errands and bookings poorly, and users cannot watch its browser live. [1] [9] [11] [12]
The model never sees the user's passwords. A separate classifier screens every outgoing message and by default blocks card numbers and government IDs. Duet needs user approval before it sends, pays, signs or deletes anything. It can make purchases through Stripe Link or Shop Pay after approval. Securities trades and transfers between accounts are blocked entirely. [11]
Duet is the fourth always-on agent from a major lab in about ten weeks, after xAI's Grok Bot, Meta's Muse and OpenAI's Dots. All four pair a dedicated cloud computer with a background agent. Duet's main differences are more connectors than Dots, which lists 4,000+, and its use of existing Claude memory and skills. [1] [2] [7] [8] [12]
The preview has no separate fee, but background checks use the plan's normal usage limits. Team and Enterprise plans are excluded, and there is no API yet. For now, Duet suits individuals and small firms more than large organizations. [12]
Launch coverage documents Duet's features, limits and safeguards, but no evaluation method, benchmark or cost-per-task data exists.
Read the full assessment
Small teams already using Claude connectors could try low-stakes reporting and drafting, accepting alerts that lag by hours and frequent approval steps.
Executive brief
Claude Duet is Anthropic's agent that works without being asked. It sends a morning brief, checks the user's connected apps every few hours, does work on a private cloud computer and asks for approval before acting. It is the fourth agent of this kind from a major AI lab in about ten weeks, after xAI's Grok Bot, Meta's Muse and OpenAI's Dots. It runs on Claude Opus 5.5, works with 6,000+ connectors, and can buy things through Stripe Link or Shop Pay once the user approves. Access is a US waitlist for Pro and Max subscribers (TechCrunch). No independent hands-on review has been published yet.
What changed and event timeline
Meta launches Muse
Muse is free with a Meta account. Reports say it reached 730,000 downloads in five days and became the top US iOS app, which pushed rivals to respond ().
Opus 5.5 ships
The model ID is
claude-opus-5-5, priced at $4/$20 per million input/output tokens, and a system card came out the same day. Anthropic says it matches Fable 5.1 on most tasks at about 40% lower cost ().OpenAI unveils Dots
Announced at DevDay, Dots are persistent agents on GPT-6 Astra. Each runs on its own cloud computer and connects to 4,000+ apps (;).
Anthropic announces Duet
It launched as a waitlisted research preview for US Pro and Max subscribers. Most outlets give Oct 21 as the date; some list Oct 22 (;).
TechCrunch frames Duet as a work tool
Anthropic says Duet is weak at errands and ticket booking and cannot show its browser live. It is aimed at briefings, dashboards and suggested actions ().
Capabilities and access
- Model: Claude Opus 5.5, running inside a virtual machine with its own browser. The VM keeps working when the user's devices are off (TestingCatalog).
- What it does: morning briefs, checks every few hours, invoice and payroll prep, contract drafts, and even web-app builds (AI Weekly; MacRumors).
- Setup: memory, skills and connected apps carry over from the user's Claude account. Everything happens in one ongoing thread.
Read the full section
- Model: Claude Opus 5.5, running inside a virtual machine with its own browser. The VM keeps working when the user's devices are off (TestingCatalog).
- What it does: morning briefs, checks every few hours, invoice and payroll prep, contract drafts, and even web-app builds (AI Weekly; MacRumors).
- Setup: memory, skills and connected apps carry over from the user's Claude account. Everything happens in one ongoing thread.
- Access: iOS, Android and the web, separate from the main Claude app during the preview. It is not available on Team or Enterprise plans, has no separate fee, and background checks use the plan's normal usage limits (VentureBeat).
Technical analysis for researchers and developers
- Credential isolation: this follows Anthropic's Managed Agents "brain vs. hands" design, so the model never sees passwords.
- Outbound filter: a separate classifier plus a safety prompt checks each outgoing message.
- Hard approval gates: the user must approve before Duet sends a message, pays, signs a document or deletes anything.
Read the full section
- Credential isolation: this follows Anthropic's Managed Agents "brain vs. hands" design, so the model never sees passwords. Action controls reuse classifiers from Claude and Claude Code (TestingCatalog).
- Outbound filter: a separate classifier plus a safety prompt checks each outgoing message. By default it blocks card numbers and government IDs, and health or personal-finance data needs explicit user instruction.
- Hard approval gates: the user must approve before Duet sends a message, pays, signs a document or deletes anything. Securities trades and transfers between accounts are blocked outright.
- Polling, not real time: checks run every few hours, so latency is measured in hours.
- Reproducibility: Anthropic has published no evaluation methodology, task benchmarks or cost-per-task figures for Duet (VentureBeat).
Claims and evidence
- Opus 5.5 was picked for its judgment about what is urgent: this is Anthropic's rationale (TestingCatalog).
- The coffee-shop demo: in Anthropic's demo, Duet flagged a rise in cancellations from 2.6% to 3.4% and traced it to a shipping change.
- Real-world use: none of the reviewed sources independently corroborates how Duet performs in practice (AI Weekly).
Read the full section
- Opus 5.5 was picked for its judgment about what is urgent: this is Anthropic's rationale (TestingCatalog). The linked video makes the same pitch: speakers praise the model's judgment on when to be proactive (01:01) and its ability to pull relevant Slack items to them proactively (01:16). These are promotional testimonials, and the captions are automatic.
- The coffee-shop demo: in Anthropic's demo, Duet flagged a rise in cancellations from 2.6% to 3.4% and traced it to a shipping change. That account comes from a Reddit summary and amounts to a staged scenario, not a measured result.
- Real-world use: none of the reviewed sources independently corroborates how Duet performs in practice (AI Weekly).
- Security testing: Anthropic reports red-teaming and an external security firm, plus a bug bounty. The findings have not been published (TestingCatalog).
Context and prior work
- A crowded field: Grok Bot (beta, Aug 11), Muse (Sep 8), Dots (Sep 29) and Duet (Oct 21) all use the same pattern: a dedicated cloud computer plus a background agent (Reddit summary).
- Muse's traction: it has reached 10.9M downloads, and Meta offers a small-business version (VentureBeat; TechCrunch).
- What sets Duet apart: mainly the more than 6,000 connectors (Dots lists 4,000+), and the fact that existing Claude memory and skills carry over.
Read the full section
- A crowded field: Grok Bot (beta, Aug 11), Muse (Sep 8), Dots (Sep 29) and Duet (Oct 21) all use the same pattern: a dedicated cloud computer plus a background agent (Reddit summary).
- Muse's traction: it has reached 10.9M downloads, and Meta offers a small-business version (VentureBeat; TechCrunch).
- What sets Duet apart: mainly the more than 6,000 connectors (Dots lists 4,000+), and the fact that existing Claude memory and skills carry over.
Limitations, safety and contested findings
- Known failure modes: Anthropic admits Duet may misread a thread or get urgency wrong (AI Weekly).
- Not built for errands: it is weak at errands and bookings, and users cannot watch its browser (TechCrunch).
- Hidden cost: background checks eat into the same usage quota as chat.
Read the full section
- Known failure modes: Anthropic admits Duet may misread a thread or get urgency wrong (AI Weekly).
- Not built for errands: it is weak at errands and bookings, and users cannot watch its browser (TechCrunch).
- Hidden cost: background checks eat into the same usage quota as chat.
- Conflicting reports:
- The launch date is given as Oct 21 or Oct 22.
- Dots' minimum price is reported as $100+ (VentureBeat) or as $20 Pro and up (Semafor).
Business and practitioner implications
- Who it suits: small firms already using Claude connectors (QuickBooks, Shopify, HubSpot, Slack) can trial it at no extra fee.
- Who has to wait: enterprises, since there are no Team or Enterprise plans and no API yet.
- Approval fatigue: gates on every message and payment make Duet safe but put a bottleneck on the human.
Read the full section
- Who it suits: small firms already using Claude connectors (QuickBooks, Shopify, HubSpot, Slack) can trial it at no extra fee. Expect quota pressure and alerts that lag by hours.
- Who has to wait: enterprises, since there are no Team or Enterprise plans and no API yet.
- Approval fatigue: gates on every message and payment make Duet safe but put a bottleneck on the human. Try it first on low-stakes workflows such as reporting and drafting.
- Developer angle: the 6,000-connector marketplace is now a channel for distribution, and integrations decide which agent fits a user's workflow.
The source trail.
Sources (12)
TechCrunch: Claude Duet app is Anthropic's bet on agents
techcrunch.comClaude Duet App: 4 AI Giants Built the Same Agent in 71 Days
Related coverage; assess separately
reddit.comClaude Duet: Claude’s New AI Agent Is Here
Related coverage; assess separately
reddit.comAI Weekly: Claude Duet waitlist
aiweekly.coLinked video: Claude Duet app is Anthropic’s bet on agents
linked video auto captions
www.youtube.com