Anthropic Takes On OpenAI, Meta With New Proactive AI Agent Claude Duet

Anthropic has unveiled Claude Duet, a proactive AI agent designed to monitor connected apps, prioritise tasks, and carry out work in the background, including when users switch off their devices.

The offering targets professionals and small business owners. Powered by Claude Opus 5.5, it is available as a research preview through a waitlist for Claude Pro and Max subscribers in the US.

Users can access Claude Duet through standalone iOS and Android apps or on Claude’s website. It is included in existing Pro and Max subscriptions but shares usage limits with other Claude products. Background checks count towards these limits, while complex assignments consume more usage, Anthropic said.

Unlike conventional AI chatbots, Claude Duet is designed to identify work that needs attention without waiting for a prompt. It sends users a daily briefing covering important items across connected apps, including emails and calendars, and provides updates when tasks are completed or require a decision.

Users can assign tasks through text or voice messages, ranging from scheduling meetings and drafting email replies to building dashboards, financial forecasts and web applications. The agent can also search past conversations and use existing Claude memories, skills and connected apps.

Claude Duet connects with more than 6,000 tools and plugins available through the Claude Marketplace, including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot and Shopify.

For instance, it can identify unpaid invoices through QuickBooks or Stripe and draft payment reminders. It can also prepare payroll information, draft job postings, create follow-up emails using customer and deal histories, and help businesses set up and manage online stores.

The agent runs on its own cloud-based virtual machine with a browser, allowing it to continue working even when users close the app or switch off their devices. Anthropic said it checks connected accounts every few hours rather than continuously, partly to limit subscription usage.

Claude Duet can also help users make business purchases through Link, Stripe’s digital wallet, and Shop Pay. Payments require user approval, while Link uses a virtual card to avoid sharing the user’s actual card number. Integration with 1Password for accessing saved website logins is also planned.

Anthropic said the agent requires approval before sending messages, making payments, signing documents or deleting information. Trading securities and transferring money between accounts remain prohibited, even with user approval. Users can disconnect apps, and only they can mark tasks as finished.

Safeguards And Early Limitations

Anthropic said Claude Duet includes safeguards designed to prevent unauthorised actions, prompt injection, and sensitive data leaks. Prompt injection involves malicious instructions hidden in content an agent encounters, such as emails or websites.

The protections include an automated system that assesses proposed actions and controls designed to keep passwords and sign-in tokens out of the agent’s direct view.

A separate message-checking system is designed to block requests to share highly sensitive information, including credit card numbers and government identification. Other sensitive information, such as health and personal finance details, can be shared only on explicit user instructions, Anthropic said.

The company said Claude Duet has undergone adversarial testing and external penetration testing, and is covered by its public bug bounty programme.

However, Anthropic acknowledged that the early version may misinterpret conversations, misjudge urgency, or miss information arriving between checks. It is currently better suited to work assignments than personal errands, such as booking tickets, and users cannot yet view its browser.

The launch adds to the growing competition in always-on AI agents that work across apps with limited human intervention. OpenAI’s Dots and Meta’s Muse target similar use cases, while Bengaluru-based Emergent’s Wingman handles tasks across email, calendars, and workplace software through WhatsApp and Telegram.

As agents gain access to more business systems, their permissions and accountability are also coming under scrutiny. Recently, CERT-In warned about unauthorised AI agent access to Indian organisations’ systems, highlighting the risks associated with deploying increasingly autonomous tools.

Anthropic said feedback from early users will inform further development. The research preview does not currently include Team and Enterprise subscribers.

The post Anthropic Takes On OpenAI, Meta With New Proactive AI Agent Claude Duet appeared first on Inc42 Media.