What makes something a real digital employee, and how to tell when you have found one.

The short version

• Google (14 October), OpenAI (12 October), Meta (29 September) and Anthropic all shipped something genuinely capable this autumn. The problem for buyers is not that they are fake, it is that the word agent now covers a chatbot and a worker with its own computer.
• A real digital employee is defined by five behaviours: it receives work, keeps going after you close your laptop, waits for other people, recovers when something fails, and reports what it did.
• The 2026 Remote Labour Index found the best agent completed 2.5% of 240 real freelance projects end-to-end, failing mostly on judgment work, not on structured follow-through.
• What separates a colleague from a demo is not intelligence. It is permissions, memory, a drawn line where a person decides, and a log you can read.
• Ten questions in any demo, starting with “what makes it start?” and “show me a failed run”, will tell you which side of that line a product sits on.

Here is what happened in the last three weeks. On 29 September, Meta launched Muse for small business, an agent that connects to QuickBooks, Shopify, Stripe and Slack, free for firms under 250 people. On 12 October, OpenAI began rolling out Dots: always-on agents, running on their own cloud computers with persistent memory. On 14 October, Google shipped Gemini 4 Argon globally, with a task mode that keeps working after you close the app. Anthropic, meanwhile, has been shipping computer-use agents, software that sees a screen and clicks it, since early 2025.

Four of the largest firms on earth, one month, one shared promise: software that does the work than answering questions about it. All of it is real, none of it is vapour, and for anyone running a small business it is the best news in a long time. It also leaves you with an entirely practical problem. Every one of those products is called an agent, and that word is doing four or five different jobs. So this is not a piece about which tool wins. It is a buying guide: here is what a genuine digital employee does, and here is how to tell, in about ten minutes of talk, whether the thing in front of you is one.

The five things a real one does

Forget the word agent for a moment. A real digital employee is defined by behaviour, and there are only five behaviours that matter. All five together make a colleague. Three out of five makes a faster way to do the same admin.

1. It receives work. Something arrives without you typing anything: an email lands, an invoice comes in, a booking is requested, a form is filled in. If you have to remember to start it, you have not hired anyone. You have bought a capable typewriter, and you are still the one who has to show up.
2. It keeps working after you close the laptop. This is the line that decides everything. Google’s task mode is the clearest example in the wild: hand over an errand, close the app, come back to a result. OpenAI’s Dots go at it from the other side, giving each agent its own computer and its own persistent memory so context survives between sessions, the thing a chatbot structurally cannot do.
3. It waits. Most working life is waiting. Waiting for the client to reply, for the payment to clear, for someone else to send the file they promised on Tuesday. It cannot be done by anything that only exists while a talk is open. A digital employee holds fifty open threads for three days and nudges the one that stalls.
4. It recovers. Things break constantly: a site is down, the file is the wrong shape, the person says no. Recovery means trying a different route, and telling you when there isn’t one, than stopping silently or producing something confidently wrong. Ask any vendor what a failed run looks like. The good ones can show you.
5. It reports. At the end of the day you get three things: what I did, what I am waiting on, and what I could not do and why. Trust is not built by a clever answer in a demo. It is built by a record you can read on a Monday morning.

Notice what is not on that list: intelligence. The models behind the October 2026 releases are not dramatically smarter than the model behind a 2024 chatbot. What changed is the harness, the tools, memory and permissions wrapped around the model, and that is engineering, not genius. Which is why the smartest model in the world is not automatically the best employee, and why any agent is only ever as useful as what you let it touch.

What October actually shipped

All four of the big releases below are genuine. They are aimed at different shapes of work, and each earns the word in a slightly different way.

The common thread is not brains. It is that each one changes the state of the world outside its own window. That is the test, and once you have it, most marketing becomes easy to read.

The line where a person decides

The design decision that turns an agent into something you can trust is where it stops. Not what it can do, what it must not do without asking. Every deployment needs one line, drawn per workflow: below it the software acts on its own, above it a person decides. Agreeing a price reduction sits above the line. Issuing a refund outside policy sits above it. Sending the routine confirmation sits comfortably below.

That is not caution for caution’s sake, it is arithmetic. An agent that is right 95% of the time but occasionally promises something outside your policy has not saved you 95% of the work. It has created a new job, checking it, and a new risk for the week no one is checking. The MIT NANDA report found the same thing from the enterprise side: the pilots that failed were rarely badly built. They were built without clear ownership, feedback loops or boundaries.

The second half of the design is recovery. What happens when it is confidently wrong, and how do you know? The answers have to be built in before it switches on: a log of every action it took, why it took it, and an off-switch that works in one move. An agent without an undo is a demo. An agent with an undo, a line and a log is infrastructure you can leave running on a Friday afternoon.

The honest limits today

None of this means these things can do everything, and you should buy with your eyes open. The most honest measurement we have is the Remote Labour Index, published in 2026 by researchers at the Center for AI Safety and Scale AI. They did not test trivia. They took 240 real freelance projects across 23 categories, roughly 6,000 hours of human work and about $140,000 of genuine client value, and handed them to agents with real briefs, real deadlines and real payment.

The results were specific. The best agent, Manus, completed 2.5% of the projects end-to-end. GPT-5, Claude Sonnet 4.5, Grok 4, the ChatGPT agent and Gemini 2.5 Pro came in between 0.8% and 2.1%. The average project was worth around $633 for roughly 29 hours of work. And the failures repeated: truncated outputs, corrupted files, floor plans that did not match their briefs, and above all, failure on the client-facing parts. The iteration on feedback. The taste. The accountability to a person expecting a result.

Read those failure modes as a shopping list, because they tell you where to aim. The agents did not fail at the structured, repetitive, multi-system work, the gathering, the chasing, the scheduling, the follow-through across three different tools. They failed at judgment. So the 2.5% is a measure of how far they are from replacing people-shaped work, and it is not remotely a measure of how far they are from carrying volume-shaped work, which they already do, all day, without complaining.

You are not hiring software to have opinions about your business. You are hiring it to carry the pile that currently stops you having any.

Ten questions to ask before you buy

Take these into any demo. They take ten minutes, and they are hard to bluff.

1. What makes it start? Name the thing that arrives which sets it working without you typing anything.
2. Where does it go when my laptop is shut? If the answer is nowhere, it is not carrying anything.
3. What does it do while it waits? Real work is mostly waiting. Can it hold fifty open threads and chase the stalled one?
4. Show me a failed run. Then show me the log for it. Vendors who have thought about failure can answer this in seconds.
5. What does it refuse to do? A good answer names things immediately, without being asked twice.
6. What can I read afterwards? Every action, with a time and a reason. Not a summary, a record.
7. How do I stop it, and does stopping lose work? One action, resumable.
8. Which of our systems does it touch? The ones you run the business on, not a new dashboard to visit.
9. Who owns it when it is wrong? A service you can call, or a licence and a support form.
10. What did it do last week that no one had to? Ask for one real example from a real customer.

Three blank answers and you are buying a demo. Ten solid ones and something quietly enormous happens: you stop checking. That is the moment people describe to us afterwards, not the cleverness, but the first Friday they close the laptop knowing the follow-ups happened, the invoice was chased, the booking was confirmed, and nothing was waiting on them.

Where Twinthos comes in

Digital employees are changing how small firms work, and the change is a kind one. The repetitive work goes to software. People get to spend their days on the parts that need a human, the judgment, the relationships, the decisions only they can make. That is the whole point of it.

This is why Twinthos builds the harness than buying the model. The intelligence is a commodity now. Google’s newest agent model costs $2 per million tokens, and it will be cheaper next quarter. What does not commoditise is the wiring: the connection to your inbox, diary and accounting; the rules taken from how your business works; the Authority Line drawn per workflow before anything switches on; and the log that proves what happened, with a switch that stops it.

A digital employee from Twinthos carries the volume-shaped work below the line and hands every judgment call above it to a person, and with a record. The model does the remembering so your team can do the thinking. If you want to know which parts of your week are volume-shaped, measured, in hours, in your business, the £999 Operations Audit is where that starts, and you keep the diagnosis whether or not you go further.

Next step

Start with the £999 Operations AuditBook a discovery call

Twinthos is a managed digital employee service, not software you install. It works across the email, diary and accounting systems your business already runs. Below the Authority Line it acts on its own; above the line a person decides.

Common questions

What makes something a real digital employee rather than a chatbot?

Behaviour, not branding. A chatbot answers in a window. A digital employee receives work without being prompted, keeps going after you close your laptop, waits for other people to reply, recovers when something fails, and reports what it did. Anything missing one of those five is still asking you to do the work.

Which of the October 2026 releases should a small business look at first?

Start from where your work already lives. Meta launched Muse for small business on 29 September 2026, it connects to QuickBooks, Shopify, Stripe and Slack and is free for companies under 250 people. Google shipped Gemini 4 Argon globally on 14 October with a task mode that keeps working after you close the app. OpenAI began rolling out Dots on 12 October: always-on agents on their own cloud computers with persistent memory. Anthropic has shipped computer-use agents since early 2025. Pick the one that touches the systems your week already runs on.

Can a digital employee really work while my computer is switched off?

Yes, and that is the whole point. Google’s task mode takes an errand and finishes it after you leave the app. OpenAI’s Dots give each agent its own cloud computer and persistent memory, so context survives between sessions, the thing a chatbot structurally cannot do. If a tool stops the moment your browser closes, it is not carrying anything.

How capable are these agents today, honestly?

The most honest measurement is the 2026 Remote Labour Index from the Center for AI Safety and Scale AI: 240 real freelance projects across 23 categories, roughly 6,000 hours of human work and about $140,000 of client value. The best agent, Manus, completed 2.5% end-to-end; GPT-5, Claude Sonnet 4.5, Grok 4, the ChatGPT agent and Gemini 2.5 Pro landed between 0.8% and 2.1%. That low number measures how far they are from people-shaped judgment work, not from the volume-shaped follow-through they already do well.

What should a digital employee refuse to do on its own?

Anything that needs judgment about your business. Agreeing a price reduction, issuing a refund outside policy or committing to a deadline sits above the line, where a person decides. Sending a routine confirmation, chasing an invoice or updating a record sits below it, where the software acts alone. Every deployment needs that line drawn per workflow before it is switched on.

Where do I start if I want one of these working in my business?

Measure first. Work out which parts of your team’s week are volume-shaped, the chasing, copying, scheduling and waiting, and how many hours they cost. The £999 Operations Audit does exactly that in your business, in hours, and you keep the diagnosis whether or not you go further with us.

Sources

1. AI agent news round-up, AI Agent Store, week of 14 October 2026 Primary report of Gemini 4 Argon, OpenAI Dots, Workday Sana and the week’s other agent releases, with pricing.
2. Meta introduces Muse for small business, Meta Newsroom, 29 September 2026 Primary source for Meta’s agent, its QuickBooks/Shopify/Stripe integrations and free tier.
3. Remote Labour Index coverage, BusinessDay, 2026 The 240-project, 23-category, $140,000-value benchmark; the 2.5% best-agent result and the documented failure modes.
4. Claude 3.5 models and computer use, Anthropic, 2024 Primary source for Anthropic’s computer-use agent capability.