What are ‘always-on’ agents, which continue to run after you’ve switched off your computer?

The most interesting development of the week in AI wasn’t a machine answering a difficult question, but a machine that no longer waits for the question. On 29 September, OpenAI unveiled Dots, a new class of agents designed to operate continuously. They have their own computer in the cloud, can connect to applications authorised by the user, move between ChatGPT, Slack and Teams, remember objectives and preferences, and continue working on a project whilst the person who commissioned them is doing something else. OpenAI says they can connect to more than four thousand applications and describes the product with a telling phrase: ‘always-on’. Until yesterday, the dominant paradigm was that of the oracle: open a window, ask a question, get an answer.

Then came the co-pilot: it stays by your side whilst you write, programme or prepare a presentation. The permanent agent attempts to take a third leap: it is no longer confined solely to the moment in which you are working. It enters the time of work itself. An example recounted by OpenAI is almost trivial, and precisely for this reason, instructive. A user had forgotten to invoice a publication for a piece of work. Their agent noticed this, prepared the invoice and sent it once it had been approved.

On the same day, another story emerged that helps to illustrate the direction things are taking. Millennium Management, a major American hedge fund, has already activated around 1,600 ‘Digital Twins’ and is making them available to over 7,000 employees. Each twin has its own identity, permissions and email address; it can carry out research, prepare for meetings and manage emails, but cannot make decisions on the employee’s behalf. According to the company, in the first pilot group, 97 per cent of users utilised the system every day.

We keep asking ourselves which jobs artificial intelligence will ‘replace’, as if the future were to take the form of a list of professions that have been done away with. In the short term, it is perhaps more interesting to observe something less dramatic: the breakdown of work into tasks that can be delegated. The roles of managers, journalists, lawyers, researchers and salespeople are full of tasks that do not require genius but do require consistency. It is invisible and time-consuming work.

The advantage of a human agent, then, is not merely ‘getting things done faster’. It is maintaining the thread of the conversation. A good human assistant is not useful because they know more than Google, but because they know that a particular negotiation is sensitive, that a certain figure needs updating every Friday, and that a promise made three weeks ago must not be forgotten. The challenge for agents is to bring this operational memory into the software. This is where the hard part begins. A chatbot that gives the wrong answer is a nuisance; an agent who takes the wrong action can cause damage. Greater autonomy means a greater need for boundaries. It is no coincidence that OpenAI has built Dots around permissions, rules and approvals: proactive searching within linked apps is read-only; the user can determine which actions are automatic and which require consent.

This revolution could be of particular help to smaller organisations. A large company can already afford assistants, junior analysts and staff dedicated to coordination. A freelancer, a small business or a ten-person editorial team often cannot. An agent who manages three projects, updates an archive, prepares materials and flags up what deserves attention does not necessarily replace an employee: they can provide those who would never have been able to hire one with organisational capabilities that were previously out of reach.

If every worker has an army of agents, however, companies could simply raise the bar: we won’t work any less, we’ll produce twice as much. The technology that promised to free us from email could end up giving us more email, generated by agents writing to other agents. That is why the crucial question is not how intelligent AI will be, but what we will do with the time it frees up. If we delegate repetitive work to machines only to fill every minute freed up with more repetitive work, we will have created a lightning-fast bureaucracy. If, on the other hand, we use that delegation to redirect time towards decision-making, relationships, creativity and responsibility, the outcome could be very different.

Finally, there is an interesting European detail. At launch, Dots is not available to Pro users in the European Economic Area, Switzerland or the United Kingdom; for Business Premium users, however, it is available in supported markets.

For years, we have imagined AI as someone we could ask: ‘Can you help me do this?’. The past week suggests a new question: ‘Can you sort this out and get back to me when it’s really needed?’. It seems like a minor difference. At work, it could be huge.