Anthropic has launched a research preview of Claude Duet, a proactive AI agent powered by its new Opus 5.5 model. Rather than waiting for a user to open a chat window and type a prompt, Duet operates in the background. It executes multi-step tasks, monitors connected software, and initiates contact when a job is done or requires authorization.
Moving Beyond the Chat Box
Currently available as a standalone app for iOS and Android and via the web, Claude Duet shifts the AI paradigm from reactive assistant to asynchronous worker. It relies on a single continuous thread that retains a user's memory, skills, and connected apps—working with more than 6,000 connectors and plugins.
The primary interface is a daily morning brief. Duet reviews connected systems and surfaces immediate priorities: an expiring job offer, redlined contract terms, or customer refund requests. Once the user addresses the morning items, the agent goes quiet and continues working on larger assigned tasks, such as compiling financial models or drafting slide decks, until it hits a stopping point.
Built-In Friction for Safety
Autonomy introduces risk, especially when an AI has access to company financials and communication channels. Anthropic addresses this by running Duet’s logins and safety checks on an isolated cloud computer.
More importantly, the system is designed with mandatory human-in-the-loop approvals for sensitive actions. The agent cannot act as a rogue financial actor. If a task requires paying a $640 county health permit fee or sending an email with a data room link to external investors, Duet prepares the action, surfaces the exact cost or drafted text, and waits for the user to explicitly click an "Approve" or "Send" button.
What This Means for SMB Operations
For small and mid-sized businesses, operational automation has historically meant building rigid logic paths. Standard automation tools move data from Point A to Point B based on strict triggers. If a step breaks or a variable changes, the workflow fails.
Claude Duet points to a more resilient automation model for unstructured administrative work. Instead of mapping out a complex flowchart to identify unpaid invoices, a business owner can simply instruct the agent to find outstanding balances and draft reminders.
The preview demonstrates a high level of operational context. In one example, the agent monitors a new location opening, tracking a 41-item checklist and proactively flagging that an espresso machine installation is slipping behind schedule. In another, it pulls data from Shopify, a CRM, and a support inbox to build an automated daily dashboard, while proactively flagging an uptick in customer churn linked to a recent shipping price change.
This level of cross-referencing previously required dedicated data analysts. Now, it can be handled by a background agent that synthesizes the data and drafts a proposed solution for a human manager to approve.
The Current Limits of the Preview
Despite the clear utility, widespread adoption is not immediate. Claude Duet is currently a research preview restricted to a waitlist for Claude Pro and Max subscribers in the US. Notably, Team and Enterprise plans do not yet have access.
Additionally, proactive agents are computationally expensive. Duet shares an account’s existing usage limits but consumes them much faster. Background monitoring and complex task execution—like rebuilding a 12-slide pitch deck with Q3 numbers—require significantly more model turns than a standard text query.
While not yet ready for company-wide deployment, Duet is a clear indicator of where business automation is heading: away from static scripts and toward persistent, judgment-capable agents that manage the mundane and escalate the critical.