While AI further integrates into our lives, it also slowly but surely changes shape. No longer are we limited to the boundaries of a textbox. Instead, we see a wave of AI agents coming our way. But what impact will this have on our organisations, and how can you keep this band of agents to play in tune?

What sorcery is this?!

Sadly, there’s no definition for agents - yet - that everyone agrees upon. At this point in time, agents and especially AI agents can mean a whole lot. And like most terms, it evolves over time. Just like a car used to refer to a horse-drawn, wheeled vehicle, then people started saying horseless car and, now, it can refer to an autonomous vehicle. Similarly, what we used to call an agent yesterday will differ from what we’ll call an agent tomorrow.

Currently, most agree that agents are computer programs that can act autonomously, use tools if needed, and “reason” in loops. They can even use tools where useful. For instance, they can search something on the internet or request information from another system. They reason by running in loops, asking themselves something along the lines of “Given my task, what I've already done and the tools I have access to, what is the best next step I can take”? Unlike traditional software that follows pre-defined rules, AI agents can learn from interactions and adapt their behaviour over time. This separates them from chatbots that just react to your input or prompt and don’t learn or reason.

Like with many other technologies (looking at you, Metaverse), the Agent hype train is omnipresent and deafeningly loud. But don’t discard agents as just another craze. It will impact every business we know today. First gradually, then suddenly. Agents will be useful in every industry but there are certain areas where it’s easier to imagine how they can help in the short term.

Take, for example, customer support. It’s costly and difficult to respond to customer requests 24/7. Agents are better suited to help out than the more static chatbots which are currently often used to replace humans, because they can respond to situations they weren’t specifically programmed for. However, aside from these more obvious examples, agents can be used for a wide variety of tasks. They could for example help follow-up on legislation and proactively suggest changes to a portfolio of legal contracts. They’re also not limited to the professional space: a Home agent could help you run your house more efficiently by monitoring your energy usage or negotiating a new contract with your energy supplier.

One thing we know for sure is that, as the underlying technology advances, agents will become more autonomous and be able to act in a broader range of activities. We’ll gradually entrust them with increasingly important tasks, and over time, there may be instances where we rely on them for certain responsibilities even more than we would a human. Just like we now trust computers more than we would trust mathematicians with performing complex calculations. However, we'll also need to be mindful of their limitations – for instance, language models can sometimes hallucinate or generate false information, so our trust will need to be calibrated to each specific capability.

Level two of five

Agents are not a new concept. They’ve already come a long way. In the late 50s, in the early days of AI, researchers had already theorised the idea of agents. In the late 60s, ELIZA was a natural language processing computer program that appeared smart: it imitated a psychotherapist and simply rephrased the "patient"'s replies as questions. It functioned much like  a chatbot, but there was no real intelligence there. Even though those interacting with the program did have a tendency to project human traits onto it, a phenomenon which later was coined the ELIZA effect.

ELIZA showed a glimpse of what would be possible but it took decades of research and development to finally bring us ChatGPT 3. In the meantime, there were mostly rule-based chatbots that weren’t all that great. Today we have a wide range of foundation models and, subsequently, GenAI tools that can show some agentic behaviour: they can do things without being explicitly asked for it.

So why are we here and is this as far as we can go? No, this is certainly not the end. But it is, perhaps, the end of the beginning. According to OpenAI’s Sam Altman (or, more correctly put, Sam Altman’s OpenAI) we’ve just entered the second of five levels: the level of reasoners. Level 1 were the chatbots, with GPT3 being the primus inter pares. Where chatbots appear smarter than they are, Reasoners bring some actual smarts to the table, being able to reason through complex tasks and solve harder problems than previous models in science, coding, and math. Sam qualifies Agents at level 3, Innovators at level 4 and Organisations at level 5. To move to level 3, Reasoners need to be able to work fully autonomously for a prolonged period in time. Level 4 requires Agents to come up with radically new ideas that go well beyond what they’re trained on. The last level would require innovators to perform the same work as an entire organisation.

So, after decades of research by some of the brightest minds in human history, we’ve only just reached level 2 out of 5? Those are rookie numbers! How come we’re not much further, you may think? There are both technological as well as societal reasons.

The technological limitations stem from two coupled factors: our available computing power and our fundamental neural network designs. We face a dual possibility: either breakthrough architectures could unlock greater capabilities even with current computational resources, or scaling up computing power with existing designs could yield more advanced systems. Recent industry developments suggest a third possibility: that current approaches may be reaching a plateau, with major AI labs reporting diminishing returns from simply making models larger. The optimal path forward may lie in architectural innovations, continued scaling, or perhaps in combining multiple approaches to overcome these limitations.

But there are also important societal reasons. Agents can challenge our legal and social frameworks in complex ways. Consider liability issues: if an agent managing your smart home's energy system makes a decision that leads to equipment damage, who bears responsibility? Or if an agent handling sensitive business negotiations misinterprets context and makes commitments beyond its authority, how do we determine accountability? Our current legal frameworks aren't equipped to handle autonomous AI agents acting on behalf of individuals or organisations. Until we develop robust governance structures and liability frameworks, many potential agent applications will remain theoretical rather than practical. Not today, but perhaps tomorrow, with the right safeguards in place.

We are the robots

As is often the case when you want to predict the future, we need to look at the past to see some recurring patterns. What agents are to the digital world, robots are to our physical world. The Volvo Cars Ghent factory introduced their first robot in the early ‘60s. This robot, the Unimate, was a curiosity and quite dangerous: people were not allowed to come close to it. It was single purpose and, frankly, it didn’t perform that one task very well either.

Its biggest achievement was to allow people to receive a glimpse of the potential of robots. However, it wasn’t making a significant difference in the work because the process of manufacturing a car was too complex for a simple robot. So what do you do when you actually have a complex task but your robot can only do one thing? You break up the complex objective into smaller, simpler tasks and have a bunch of single-purpose robots take up each of those smaller functions. That’s the logic of a manufacturing line, and that’s how it works with agents, too. Currently, agents are very limited in what they can do, so we use multi-agent systems to tackle more complex tasks. For each small task, we have a specific agent and we let the agents figure out how to proceed. The short-term future will be multi-agent oriented.

This gradual introduction of robots in manufacturing hasn’t reached an endpoint, yet, as some had predicted. There are still 1000s of highly-skilled workers who help construct cars alongside robots. Even Elon’s robot-heavy megafactories run on lots of very experienced and talented people to make sure every car is as good as it can be.

For some of these tasks we did pass an important inflection point, where we can’t imagine humans doing a certain task anymore. In car manufacturing, spot welding frames is almost exclusively done by robots (the precision brings more reliable road safety), as well as painting (more reliable paint jobs with a minimal amount of paint). For agents, we’ll have to see which activities reach this inflection point first. Data-heavy tasks with lots of analysis involved first come to mind.

Seize today, shape tomorrow

Nobody can predict what awaits us but it’s important that we work towards a desirable future. Think about the first robot that entered Volvo Cars Ghent. Which first agent will you install in your organisation? Whatever task it will work on, don’t let it be a gimmick. It has to do actual work, just like that first robot. You’ll need to carefully monitor it, let everyone get used to the idea, and inspire people to think of what the actual shift will be. It won’t be that first agent that will make the big difference, but the hundreds of agents that come after it. Just like in manufacturing you’ll need to gradually introduce them.

As we get used to having agents do simple tasks, we need to learn to ask the following question: “What would an agent do?” Don’t think about how you can replace certain tasks a human currently does by an agent. We need a native, agent-first approach. 

Imagine a typical customer support flow where questions and feedback come into a ticket system. The first phase might be to replicate human behaviour and have one agent handle ticket per ticket. But a more agentic approach would be to take in all recent customer support tickets at once and treat them as a whole. The context window of an agent can be much bigger than that of a human. For an agent it makes no sense to treat everything one by one, especially as multiple requests may signal a bigger issue or outage and could have an effect on how you respond.

So that would be our advice to you today: roll out your first agent in the immediate future, learn from it and shape the agentic way of solving tasks tomorrow.