Anthropic launches Claude Duet background AI agent for US subscribers

A proactive shift in AI

Anthropic recently introduced Claude Duet, a new type of AI agent designed to work in the background rather than waiting for a user to type a specific prompt. Unlike standard chatbots that require a constant back-and-forth conversation, this tool is built to be proactive. It monitors connected applications and only reaches out to the user when a specific action or piece of information requires human attention.

The company is positioning this tool specifically for professionals and small business owners who manage many different projects simultaneously. By acting as a background assistant, it aims to reduce the cognitive load of checking multiple apps for updates. The goal is to move the AI from a tool you use to a partner that works alongside you throughout the day.

As gHacks reported, the service is currently being released as a research preview. This means the company is looking for feedback on how the agent handles real-world workflows before a wider release. Users can currently access it on the web, as well as on Android and iOS devices.

How the background agent works

Claude Duet operates inside its own virtual machine, which allows it to continue running tasks even when a user is away from their computer or has closed their chat window. This architecture is what enables the agent to perform long-running jobs without requiring a constant active connection. It is powered by the Claude Opus 5.5 model, which provides the reasoning capabilities needed for complex tasks.

Every morning, the agent provides a summary of important items from the various apps a user has connected to the system. Throughout the rest of the day, it continues to monitor those same sources for changes. If it detects something that needs a response or a decision, it sends an alert to the user. This allows the agent to filter out noise and only highlight what is necessary.

Users can also assign specific jobs to the agent to complete autonomously. Anthropic lists several examples of these tasks, including building dashboards, preparing financial and sales forecasts, and prototyping or developing web applications. It can also handle logistical tasks like scheduling meetings across different calendars.

Privacy and security measures

Because an agent with this level of access can see work messages and other sensitive data, Anthropic has implemented specific safeguards. The company states that Claude Duet does not have direct access to sign-in tokens or passwords. This is intended to ensure that the agent can interact with the apps without having the ability to compromise the user's primary credentials.

The system is also designed to recognize and protect highly sensitive information. Anthropic claims that safeguards are in place to prevent the agent from sharing data like government IDs or credit card numbers. The company also conducted security testing on the system to ensure these boundaries remain intact during operation.

Despite these protections, the agent is still designed to judge which events deserve a human's attention. It tries to determine when it should act on its own and when it must ask for permission first. This balance is intended to give the user control over the agent's autonomy while still allowing it to handle the heavy lifting of monitoring data.

Subscription and usage limits

Claude Duet is not available for users on the free tier of the service. To join the research preview, users must have a paid Claude Pro or Claude Max subscription. Additionally, the preview is currently restricted to users located in the United States who sign up for a specific waitlist.

Once a user is granted access, they must manually connect the apps they want the agent to monitor. This allows the agent to begin generating the morning rundowns and sending daytime alerts. However, users should be mindful of how the agent consumes their account's resources. Background work performed by the agent counts toward the same usage limits as regular chat interactions.

For heavy users, this shared limit is an important consideration because the agent can perform many tasks while the user is offline. If the agent performs a large number of actions, it could potentially exhaust the user's available capacity for the day. Anthropic is being transparent about these limits as they gather data on how the agent behaves in a production environment.

Current limitations and future

Anthropic has been open about the fact that Claude Duet is still in its early stages and has some rough edges. For example, the company admits that the agent can sometimes be overly enthusiastic in its actions. They also note that the agent is not currently the best fit for personal tasks, such as booking travel tickets, which may require more nuanced human interaction.

One technical limitation is that Claude Duet cannot currently show the user its browser window. This means users cannot see exactly what the agent is doing behind the scenes as it navigates apps or builds dashboards. This lack of visibility makes it harder for users to audit the agent's step-by-step process in real-time.

The launch places Anthropic in competition with other companies like Meta and OpenAI, which are also developing assistants that act rather than just answer questions. Anthropic has not yet announced a timeline for when the research preview will end or when the service will expand beyond the United States. For now, it remains a specialized tool for a small group of early adopters.