Personal AI agents are going mainstream. We have run one on a Mac mini since March

This week Anthropic launched Claude Duet, an app where Claude messages you first. It reads your email and calendar, writes you a morning brief, checks your apps every few hours and keeps working after you put your phone down. My first reaction was a laugh, because that is what our own AI operator has done for my businesses since March 2026, from a Mac mini on a desk in Scotland.

That is not a boast. It is the point of this article. Until now, an always-on agent was something builders put together themselves. Over the last three months four of the biggest AI companies have each shipped one as an app. The rest of the market is about to find out what the builders already know: the benefits are real, and so are the rough edges.

Four launches in eleven weeks

Between August and October 2026, every major lab released a version of the same idea.

• Grok Bot (xAI, 11 August). "AI teammates" with their own cloud computer that sign into your tools, work across apps and come back only when something needs approval. You can run several at once, with one acting as chief of staff over the others. In beta for SuperGrok and Cursor subscribers.
• Meta Muse (September). A personal agent you reach through its own app or WhatsApp, aimed at everyday tasks, purchases and travel. It is free to start, and Sensor Tower estimates put it at 10.9 million downloads by 20 October.
• OpenAI Dots (late September). Agents that investigate issues, build apps and move projects forward between conversations, working through ChatGPT, Slack and Teams. Available on Pro and Business Premium, with an Enterprise beta.
• Claude Duet (Anthropic, 21 October). A separate app built on Claude Opus 5.5, aimed squarely at professionals and small business owners. Morning brief, background checks, larger jobs handed back as files, and approval before anything is sent, paid, signed or deleted.

The details differ. The shape is the same: an agent with its own computer, access to your apps, memory of how you work, and a rule that it checks with you before it does anything you cannot undo.

What we built before the apps existed

Our operator runs on an always-on Apple Mac mini. It talks to me on Telegram, the way you would message a colleague. Here is what it actually does each day:

• A morning report at about 5am. Revenue, website traffic, ad spend and cost per sale across our businesses, plus anything that broke overnight.
• Scans through the day. It checks the pipeline, the inboxes and the scheduled jobs, and tells me only what needs me.
• Real work on request. It researches prospects, drafts proposals and emails, builds reports, updates the CRM and writes articles like this first draft.
• Memory that survives restarts. Notes on every decision, every rule I have set and every mistake it has made, so it does not repeat them.

It is the same idea as Duet. The difference is that it is built around our own tools, our own rules and our own hardware, and we have had seven months to learn where it goes wrong.

What seven months taught us that the launch videos do not show

Approval gates are the product. Every one of these launches stresses that the agent asks before it acts. That matches our experience exactly. Ours can draft a customer reply in seconds, but it has never sent one to a customer without my say-so, and that rule is written down where it cannot drift. The value is in the drafting and the noticing. The risk is in the sending.

An agent that reports confidently can still be wrong. In July ours made several reporting errors in a few days. In one case it saw a file in a folder and reported that an email had gone out. Something being there is not the same as it having happened. We fixed it with a hard rule: before any claim reaches me, it must check the thing that proves it, such as the send log or the status field, and say what it checked. Anthropic admits Duet "might misread a thread or miscalibrate what's urgent." Plan for that from day one.

Memory needs maintenance. An agent that remembers everything also remembers things that are no longer true. A good part of the effort is keeping its notes current, pruning old rules and making sure the important ones are read every time, not just when it happens to look.

Background work costs money. Duet's own help page says its background checks count against your usage limits, which is why it checks every few hours rather than constantly. We hit the same trade-off. Every scan, brief and check-in has a cost, so you decide what is worth watching.

Your systems decide how useful it is. An agent is only as good as what it can reach. Connectors make the common tools easy. The value for a real business is usually in the awkward ones: the accounting package, the industry system with no clean connector, the spreadsheet everyone relies on.

Why the apps will not fit every UK business yet

The consumer apps are a big step, and they will get better quickly. For a UK small business today there are practical limits worth knowing:

• Availability. Claude Duet is US only during its preview, on Pro and Max plans, through a waitlist. It is not on Team or Enterprise plans yet.
• Microsoft 365. Duet does not support it yet. A large share of UK SMEs run their email, files and calendars on it.
• Your rules, not theirs. A consumer app sets its own approval rules. A business usually needs its own: who can approve a payment, which customers get a personal reply, what must never be automated.
• Where your data sits. These agents run on the provider's infrastructure. For some businesses, particularly in regulated work, that needs a proper look before anything is connected.

None of that is a reason to ignore them. It is a reason to be clear about what you would hand an agent before you sign up for one.

What we think the personal agent world looks like

Here is our view, based on running one every day. Treat it as opinion, not prediction.

1. Everyone gets a chief of staff. The first job these agents do well is triage: reading everything, deciding what matters and getting the first draft started. Within a couple of years an owner-led business without one will feel like a business without email.
2. Small teams of agents, not one. Grok Bot already lets one agent manage others. The pattern we expect is a lead agent that knows the business, with specialists for the inbox, the books, sales follow-up and reporting.
3. The job moves from doing to approving. Owners will spend less time doing tasks and more time making the calls an agent brings them. The skill that matters is writing good rules about what needs you.
4. The winners connect to the boring systems. Every agent can read Gmail. The ones that pay for themselves will be the ones wired into your CRM, your accounts and your industry software.
5. Trust is earned slowly. Agents will be given more freedom over time, but only after a track record. Businesses that start now, with tight approval rules, will be the ones that can safely hand over more later.

What to do now

You do not need to wait for an app to reach the UK. Start with three questions:

1. What would you hand an agent first? The morning brief, inbox triage and invoice chasing are good starting points.
2. What must it never do without asking? Write the list down.
3. Where does your work actually live? If it is spread across Microsoft 365, a CRM and a few awkward systems, that shapes what will work.

That is the conversation we have on an AI Operations Day. We map where an agent would save you time, what it needs access to, and which rules keep it safe, then build it around your systems rather than ours. If you want the background first, see how we approach AI employees, Claude for business and managed operations.

The builders have had this for a while. Now everyone will. The businesses that get the most from it will be the ones that thought about the rules before they handed over the keys.