The emergence of AI agents will democratize access to a resource that has historically been the preserve of the rich: the service of others
When we hear the word ‘agent’, many of us will have in our mind’s eye an image of a sharp-suited pistol-carrying man, turning towards us from the silver screen. A shot rings out. Blood trickles down the scree and a guitar twangs with that iconic riff.
Soon, however, we will become familiar with a very different type of agent – the AI-based software agent. These agents will be less like James Bond and more like Miss Moneypenny, the personal assistant who booked Bond’s travel and took care of all of the administrative tasks that Bond was too busy to tackle.
AI-based software agents are the next evolutionary step in the development of computing. They will push us toward a world where a computer is less and less a hand tool that helps us do something, but rather a power tool that does things for us.
AI agents are currently being developed in software labs around the world, and are set to become pervasive in 2025 and beyond. They will fundamentally reshape how IT is created, deployed and optimized – and will further reshape how many (if not, in time, all) work processes are undertaken.
From do-it-yourself to do-it-for-me
An AI agent (AIA) is, as Amazon Web Services defines it, a software program that performs self-determined tasks to meet predetermined outcomes. A human sets the goal, but then the AIA completes the tasks necessary to achieve that goal without human intervention. As these tasks become more complex, the software is increasingly executing the task, rather than just helping a person to execute the task.
Consider a customer service interaction. Currently, a human call center agent sits in front of a screen and talks with a customer over the phone, accessing a wide array of different information, such as the customer’s identification details, account details, and their previous interaction history. In real-time, the human agent is tasked with resolving the customer’s issue, making sense of all the information they have to hand.
In the future, the human-set task will be “resolve the customer’s problem to their satisfaction,” and the AIA will do the rest. Through a natural language interface – and increasingly, through an AI-generated avatar (that is, a digitally-rendered human) – the AIA will undertake the task at a pace and with a degree of accuracy that a typical human agent finds hard to reach.
Sierra’s Agent OS is a good example of how the inherently problematic domain of customer service will be re-engineered by very smart software. A relatively new startup founded by Bret Taylor, the former co-CEO of Salesforce, and Clay Bavor, who formerly led Google Labs, Sierra has built an autonomous agent that is able to operate complex, multi-step workflows, and securely access and integrate systems (internal and external, where required) to gather data and take action. It can self-supervise too, ensuring the customer interaction stays ‘on brand’ and ‘on message.’ At a fraction of the cost of employing a real person, and with none of the human problems that humans bring to work, such AIAs will soon be commonplace.
AI agents are already beginning to appear in a variety of other areas far from customer service. One of the most important and interesting is in software development itself. As prominent venture capitalist Marc Andreessen said in 2011, “Software is eating the world.” Now, software is beginning to eat software as coding becomes agent-driven. New generations of software languages and libraries are “abstracting” technical complexity (how to build) and replacing it with objective complexity (what to build).
A tool such as Devin, developed by Cognition, is accessible to even non-technical people. The process starts with outlining an objective, either by typing or talking. For example: “Create a 3D dashboard that shows Days Sales Outstanding for the South East region, plotted against average daily temperatures. Then put it in a presentation template and bullet point the important analysis about it. Then send it to the CFO via email from me, and ask her for comments and/or approval.” The AIA will complete the task, with no human software engineer involved at all.
Devin uses multiple technical tools – including browsers, large language models (LLMs), code repositories and open source libraries – to evaluate, reason, decide and act without the need for the constant prompting that existing LLMs such as GPT3 or Llama require. The task setter – the human outlining the objective – can, if they want, observe the AIA in action and make adjustments to the objective, but this requires a deliberate slowing down of the execution. The above example is completed in seconds, not minutes, let alone the days and weeks a normal software development shop would typically require to complete such a task.
Devin and other emerging breakthrough agents such as Replit or Talent Agent, and ones from established software vendors such as Salesforce, are an extension of the functionality that software commonly has now. Developing a software product or program requires the execution of millions of tasks specified by the human programmer but undertaken by software.
AI agents take this logic to a new and unprecedented level. They are more than just autocomplete tools or AI assistant coding tools, such as Tabnine, which already have millions of users. Agents like Devin promise to remove the technical ‘human in the loop’ and replace them with a non-technical human able to ask for things in the language of the enterprise they’re engaged in, not through an intermediary that knows the foreign language of technology.
The new computer language: English
This marks the end of an era in which people had to know how to talk to computers through computer languages. It is a huge breakthrough – one that creates huge potential for technology to be of even greater service to humans than has hitherto been the case.
Imagine a world in which all of us can have our own Miss (or Mr) Moneypenny: someone who’ll book our travel, make our hotel reservations, file our expenses, apply for our car insurance, or make the kids’ dentist appointments (and maybe even renew our license to kill). We might have an agent – perhaps multiple specialized agents – at our beck and call 24/7, 365 days a year. An AIA will never phone in sick or unexpectedly need to take the day off. It is able to take care of the tasks we hate doing (for me, those pesky expense reports) and frees us up to do the things we love, value and monetize.
Though this sounds like a fantasy, it is in reality the world in which the superrich have always lived. The wealthy have always had agents – slaves, serfs, household staff members, employees – who execute things on their behalf and at their command.
Once upon a time, the less wealthy had access to these people too; when you wanted to go on vacation you went to a travel agent, a real live human in a store on main street who you could talk to and say, “I want to go away for a fortnight, somewhere sunny and near the sea, that has a golf course and doesn’t allow children in.” Miraculously, the travel agent would book your trip to the Quinta do Lago Laranjal, and away you would go.
Since the aforementioned Marc Andreessen invented the browser (he was co-author of Mosaic) and brought the Internet to life, however, we have all been slowly conditioned to forgo the travel agent and search online ourselves. Wonderful as this has been, the DIY element to this has, over time, become more and more onerous and tiresome.
Searching for that perfect golf resort on Google brings up link after link. Even after refining the search terms and wading through online review after online review – which could be from a real customer, but could just as well be from a hotel employee – one is often left longing to turn to a travel expert who knows you and could just recommend the perfect place.
With this full circle moment, the agent is reborn, now as a piece of software. An AI agent will undertake the task and simply be judged on the outcome.
Of course, the issue of trust will be essential to this judgment. On the journey to agents becoming a normal part of a workflow, we will need to feel assured that AI agenda are both technically ‘competent’ – that is, accurately completing the tasks it is set – and acting in the manner of a fiduciary – with our, not its, best interests in mind.
This is not a hypothetical concern. Already, we have seen that LLMs have been prone to “hallucinations,” a polite phrase for their not-uncommon errors and inaccuracies. Reducing these errors is a priority for LLM developers. However, the fiduciary aspect is a new threshold for software to cross. The notorious “paperclip problem” of AI highlights the problem. It posits that an AI given an objective to create paperclips might attempt to turn everything into paperclips, destroying the world in the process of fulfilling its mission.
This has rightly led to material investments in creating “safe” AI. These initiatives will be fundamental to agents becoming trusted partners of humans – which, in turn, will be a key step towards their commercial adoption.
A new era of technology democratization
AI agents will see the democratization of a historically constrained resource – people who do things for you. Now you won’t need to be the senior vice president to have an assistant, or a software development team, or a horde of interns. Even the interns will have interns, in the form of their AIAs.
Admittedly, the path to AI agents has been a rocky one. The ugly history of “bots” and robotic process automation (RPA) has left many of us cynical and weary; one could be forgiven for thinking that in most Fortune 500 companies, the customer service function would be more aptly named the Department of Customer Disservice. AI agents are still in their infancy, yet the possibilities set to be unleashed by their maturation will be extraordinary.
We live in an age of incredible technology and live lives that would baffle our recent ancestors, yet software is, according to one of its primary inventors, still pretty dumb. AIAs will change that. Bill Gates, for one, believes that AI agents are about to “completely change how you use a computer.” From healthcare to education, Gates argues that AI agents will rewire how humans interact with technology. Far from technology being at a plateau, as some people argue, Gates sees a future in which software has huge room to grow, become smarter, and be more useful.
Historically, the things that humans monetize – roughly what we call work – are aggregations or disaggregations of tasks. (The event planner figures out how to do search engine optimization, but as a result becomes so successful and busy that she has to hire an SEO specialist.) AI agents are set to take us into a new era of task aggregation and disaggregation in which human work will be reimagined, as the work that agents can execute expands dramatically.
The future of human and agent work
The specter of human unemployment will loom large in the negotiation between human and AI of who does what. Pessimists point to the impact of AI on software developers that we are already seeing. Yet Gartner has argued, in an analysis released in August 2024, that AI will not actually replace software engineers: it may, in fact, require more.
Likewise, the democratization of agency unleashed by AIAs will paradoxically create the need for more human work, as agents give us the time and energy to think of new things to do – hopefully better, more satisfying, and more profitable things.
Humans and their AI agents promise a new era of old and new work in which business-as-usual will be the perfect recipe for irrelevance. Now is the time to investigate the power of agents to help you drive cost out of your process execution, and increase velocity and accuracy. AI agents will not be a secret service – rather, they are set to quickly become a source of competitive advantage that is available to all.
Ben Pring is co-author of several best-selling and award-winning books including “What to Do When Machines Do Everything”