Blog// artificial intelligence

Anthropic launches Claude Duet: Opus 5.5 agent works in background and initiates contact on its own

Anthropic launched Claude Duet, an Opus 5.5 agent with a cloud browser. It runs in the background, connects to 6,000+ services, and initiates dialogue. Research preview for Pro/Max in the US.

23-10-2026 · 7 min read
Anthropic launches Claude Duet: Opus 5.5 agent works in background and initiates contact on its own

Anthropic introduced the Claude Duet app — a personal AI agent based on the Opus 5.5 model. Unlike a typical chat, Duet performs tasks in the background, monitors events in connected services, and initiates dialogue on its own when human input is needed. The product targets professionals and small business owners. According to Habr, the agent can already prepare morning briefings, check email and payment systems, and create dashboards and web app prototypes. For business, this means the autonomous agent category is moving from labs into working tools, albeit with significant caveats.

Architecture: virtual machine and persistent context

Duet does not run inside the user's browser. The agent has its own virtual machine with a browser in the cloud. Tasks continue executing when the laptop or phone is turned off. According to Habr, memory, skills, and app connections are transferred from the main Claude account into Duet. Communication takes place in a single continuous dialogue — text or voice. When work is finished or user input is required, the agent sends a notification. This approach removes the session time limit: the agent can work in email or CRM for an hour, gathering data for a report, without requiring an open tab.

Initiative over reaction: morning briefing and scheduled checks

The main difference from classic assistants is initiative. Each morning Duet compiles a briefing of important events, then checks connected accounts every few hours. The check frequency and list of topics to monitor can be changed in a normal conversation with the agent. The publication does not specify whether there is a rigid schedule or the agent decides when to check a given source, but flexibility of configuration via natural language is claimed. For a business owner, the tool shifts from the “do what I say” category to “watch what I miss”.

Tasks and integrations: from emails to dashboards and prototypes

The list of claimed capabilities is broad: organizing meetings, preparing reports and forecasts, creating dashboards, prototypes, and full-fledged web applications. Duet has access to more than 6,000 integrations and plugins from the Claude Marketplace. Examples: finding missed emails in Gmail and messages in Slack, checking unpaid invoices in Stripe, generating client emails that account for deal history in HubSpot. This covers the routine of sales, finance, and operations. The agent does not just search for information — it executes a chain of actions: finds the debt, opens the deal card, composes a reminder, and sends it to email drafts.

Limitations: research preview, urgency errors, and closed access

Anthropic honestly states the status: research preview. The agent may misinterpret conversations and misjudge urgency. For business this is critical: a false “urgent” alert annoys, while a missed genuinely hot email costs money. Access is currently waitlist only. At launch Duet is available exclusively to Pro and Max subscribers in the US. Apps exist for iOS, Android, and a web version at claude.ai. Payment with Russian cards and access without a VPN are impossible. The publication does not specify timelines for geographic expansion or the appearance of enterprise tiers with role-based access.

Competitors: Dots from OpenAI and Grok Bot from xAI

In usage model, Duet is close to Dots from OpenAI and Grok Bot from xAI. All three products run in the cloud, use a remote browser, and connected apps. This forms a new standard: an agent as a cloud employee with access to the web and APIs. The difference so far lies in the base model (Opus 5.5 at Anthropic) and the integration ecosystem. For the user, the choice will be not between “agent” and “chat”, but between ecosystems: Anthropic, OpenAI, xAI. Lock-in to an ecosystem will begin at the level of configured integrations and familiarity with a particular agent's working style.

What this means for business in Russia and the CIS: access, alternatives, and preparation

There is no direct access to Duet right now. You need a US account, a payment card that works abroad, and a stable VPN. Pro ($20) and Max ($100) plans are billed in dollars. For testing, you can create a foreign Apple ID or Google Play account and buy the subscription through the app store if you have a linked foreign card. A ready-made solution for a Russian company today carries the risk of losing access at any moment.

What makes sense to do this week:

  1. Routine audit. List 10–15 recurring operations: checking incoming messages, gathering ad stats, reconciling payouts, preparing the weekly investor report. Mark where historical context is needed — correspondence, CRM, finance.
  2. Integration check. Open the connections list in your current AI tool (ChatGPT, Yandex GPT, GigaChat, Claude API). Is there access to your AmoCRM, Bitrix24, Yandex.Metrica, 1C, banking APIs? If not, the agent won't be able to work with local data.
  3. API pilot. Don't wait for Duet. Set up a simple scenario via your current model's API: a daily scheduled query to email or CRM with the prompt “find what requires action and send a summary to Telegram”. This delivers 80% of the initiative effect without waiting for access.
  4. Local alternatives. Watch for announcements from Yandex (Neuro, Business Assistant), Sber (GigaChat API with functions), and MTS. They have out-of-the-box access to Russian services and ruble billing.
  5. Data security. Before granting email or CRM access to any cloud agent — foreign or local — run an audit: what data leaves in the request, where logs are stored, whether legal and security have signed off. This is especially relevant in research preview.

Bottom line

Claude Duet is the first mass-market product where an AI agent runs in the cloud continuously, controls a browser, and initiates communication on its own. Technologically, this is a step from text generation to executing workflows. Practical value for Russian small business is absent today due to closed access and preview status. Strategic value lies in understanding where the market is headed: in a year or two such agents will become a standard option in enterprise tiers of major providers. The winner will not be the one who gets access to Duet first, but the one who by that time has already broken down processes into atomic tasks, set up data exports in a convenient format, and taught staff to assign tasks to the machine clearly and measurably.

Frequently Asked Questions

What is Claude Duet?

Claude Duet is an autonomous AI agent developed by Anthropic based on the Opus 5.5 model. It can work in the background, initiate dialogue on its own, and integrate with more than 6000 services.

What are the main differences between Claude Duet and other AI agents?

Key differences include autonomous operation mode, the agent's ability to initiate interaction itself, and extensive integration with external services. This enables it to proactively perform tasks and provide information.

Who is Claude Duet intended for at this stage?

At the current stage, Claude Duet is available in research preview format for users of Pro and Max plans in the United States. This allows Anthropic to collect feedback and refine functionality before a broader launch.

Based on: Habr

Analysis & review by
Alexander K.

Web development, digital advertising and marketing expert since 2012. Founder of AIFORSALE.

How we can help
Online storesMarketingAI agentsContact us
← All articles