Claude Duet, Muse and Dots: The Rise of the Background AI Agent
On 21 October, Anthropic released a research preview of Claude Duet, an AI agent built for professionals and small business owners with more to keep track of than time allows. Rather than waiting for a question, it watches the apps it’s connected to, checking them every few hours, and messages its user when something needs attention. Each morning it sends a short brief on what matters most. It can be handed tasks by text or voice, and it keeps working on its own cloud-based computer when its user steps away.
It isn’t alone. Meta’s Muse and OpenAI’s Dots, launched in September, work along similar lines: AI agents with their own computer in the cloud, able to keep working in the background. The category has moved from concept to product in a matter of weeks.
For Australian businesses, this is a preview rather than something to plan around. Duet is limited to US subscribers on Anthropic’s Pro and Max plans while it’s in research preview, with access through a waitlist and no Australian timeline published. It isn’t yet available on Team or Enterprise plans. But the question it raises doesn’t depend on the launch date: if an agent like this tried to help run your business, how much of it could it actually see?
Anthropic’s examples are deliberately ordinary. Duet can surface emails and Slack messages that slipped through, check QuickBooks or Stripe for unpaid invoices and draft reminders, prepare agreements for signature, schedule meetings and build simple dashboards or forecasts. Its listed integrations include Gmail, Google Calendar and Drive, Slack, HubSpot, DocuSign and Calendly, with Stripe, Link and Shop Pay as partners. A visual Shopify storefront editor is listed as coming soon.
The safeguards are just as telling. Duet prepares work on its own, but asks for approval before sending a message, making a payment, signing a document or deleting anything. Some actions, such as trading securities or moving money between accounts, are off-limits even with approval. Anthropic is also candid that it can misread a thread or misjudge how urgent something is. It’s a capable assistant with a short leash, which is probably the right shape for this stage.
An AI agent works in two ways. Where a system has a proper connection, such as an API or a purpose-built connector, the agent can read and act on information directly and reliably. Where it doesn’t, the agent falls back on operating a website the way a person would, through a browser, which is slower and more likely to go wrong when a page changes. Duet does both, but its browser runs out of sight, and Anthropic notes that personal errands done this way are a weaker point than work tasks.
For an eCommerce business, that distinction matters a great deal. An agent that can query your storefront, ERP and accounting system directly can answer “which wholesale orders are overdue?” or “which products are about to run out?” with confidence. One that has to click through a supplier portal, an old admin panel and a spreadsheet emailed every Friday will be guessing at best. The useful work these agents promise sits almost entirely on top of systems that are already connected and accurate.
It’s also worth noticing which tools appear on launch lists. Duet’s named integrations lean towards platforms popular with US small businesses, and Microsoft 365 isn’t yet supported. Many Australian businesses run on Xero or MYOB for accounting, Microsoft 365 for email, and an ERP such as Pronto or Dynamics 365 behind their store. Where those systems already have clean, well-documented integrations, the gap is easy to close as agent tools mature. Where data only moves through manual exports, there’s nothing for an agent to plug into.
Giving an agent access to business systems is a decision about permissions, not just convenience. Most platforms were designed around people logging in, and many small businesses share a single admin login more widely than they should. That’s a manageable risk with a handful of staff. It’s a bigger one when the user is software that runs every few hours and acts on what it finds.
The sensible approach is the same one good account security has always relied on: access tied to a specific user, limited to what the task needs, easy to revoke, and with a clear record of what was done. Duet’s design reflects some of this. It doesn’t see passwords or sign-in tokens directly, and it asks before anything irreversible. But the systems on the other end need to support it too. A platform that can’t give a read-only role to an assistant, or can’t show who changed a price and when, is harder to connect safely to anything, AI or otherwise.
There’s a second reason to pay attention. The same kind of agent that chases a business owner’s unpaid invoices could, before long, be placing a reorder for one of your B2B customers. Trade buyers are often small business owners themselves, and an agent that notices stock running low and prepares an order for approval is an obvious next step. When that happens, the suppliers whose ordering platforms are easy for software to work with, with clear account-specific pricing, accurate stock and predictable checkout, will be easier to buy from than those that rely on a phone call or a PDF price list.
None of this needs a project today. A useful first step is to list the systems your business depends on and check, for each one, whether it has a proper integration or API, whether access can be set per user, and whether the data in it is current. Then think about the routine checks your team does every week, such as overdue invoices, low stock, stuck orders or unanswered enquiries, and whether the information they rely on lives in one place or five.
Proactive agents will arrive in Australia on their own timetable, and they’ll keep changing quickly once they do. The businesses ready to use them won’t be the ones that signed up first. They’ll be the ones whose systems were already connected well enough for an agent to find the right answer.