After the initial wave of generative AI, the whole world now seems to be talking about agents
And while the words agent and agentic are now suddenly everywhere, they have a long history and well-established meanings.
Defining what you need is important
Whilst it might seem simplistic to focus on definitions, it is crucial to do so for this topic. Why? Because there are many different definitions of agents and agent systems. If we are ever to form a coherent, interconnected ecosystem of agents, we need to start with clarity and consistency around the terminology. Ironically, the terminology around agents is some of the oldest and best-established in the AI field. It is also interesting to consider people’s backgrounds and interests. People who sit more on the business side tend to define an agent based on how it affects the business whereas technical people tend to define an agent based on what it can functionally do. Here, we define key concepts to help demystify this topic for all.
An agent
An agent is any entity that works on behalf of another entity, working to accomplish high-level objectives often using specialist capabilities. Agents have the autonomy and authority to take actions that modify their world.
A key aspect of this definition is the ability to take action – literally exercising agency. AI can be a great assistant, but if it does not have the ability to take action, it cannot be an agent. An agent, however, does not necessarily have to use AI. Many non-AI systems are agents too (a smart thermostat is a simple non-AI agent). Similarly, not all AI systems are agents. So, let’s take a deeper look into what some of the key terms are around agentic systems to better understand what an agent is and is not.
.
IThese three terms could well be thought of as quite similar, however it is important to define them and discuss how they differ from each other and why it matters in this context.
- Autonomy is a measure of the degree to which an entity can independently make choices.
- Agency refers to the degree to which an entity has the capacity to act on those choices.
- Authority refers to the specific scope or limitations of the actions an entity can take.
All of these are continuous spectrums, not binary properties. The thermostat mentioned earlier has a high degree of autonomy (it can decide what needs doing without human intervention), a high degree of agency (it can take those actions without oversight), but a low degree of authority (it can only do one thing – switch the heating on or off). Using this example, you can imagine many agents and debate how much autonomy, agency, and authority they have.
One more thing to consider here is that the designers of agentic system are the ones deciding how much autonomy, agency, and authority agents have. So, we should all ask these questions when making agentic systems to ensure they can operate under the conditions they need but do not have complete freedom beyond the need for the use case. These three properties provide a useful framework to discuss and assess the risks and opportunities of agents. An agent with high autonomy, agency, and authority could be extremely powerful, but also very risky. Therefore, what is the necessary level of autonomy, agency and authority? It is up to you, as the human designing the system, to decide!
A multi-agent system (MAS)
So far, we have talked about what an agent is, but not much about what happens when you have lots of agents working together. Whilst there are lots of great use cases that require single agents, often the value will be had with a multi-agent system. But what is a multi-agent system (MAS)? Simply, we define a MAS as a system made up of multiple independent agents that operate in the same environment.
It is also worth noting that systems that use agents are sometimes called agentic architectures/ frameworks. Now that we have defined many of the key terms, let’s move on to some of the key concepts.
World models
Agents operate within a specific “world”, representing the totality of what they can sense and act upon. This could be a narrowly defined software environment, or the actual, physical world. Coming back to our thermostat example, with a limited world model, the thermostat only knows about temperature. An advanced thermostat with a richer world model might understand occupancy patterns, weather forecasts, utility pricing, and user preferences.
This comprehensive understanding allows it to make decisions that appear intelligent rather than merely reactive – turning down heating before you leave or pre-warming before expected return – building trust through apparent understanding of context. An overly simplistic world model can lead to poor performance. If a customer service agent does not have good contextual information about the customer and their situation, its advice would likely be of very low quality. World models are something that all humans have, and while they may differ slightly between people, our shared model of the world allows humans to collaborate, anticipate, and empathize with each other in order to solve tasks efficiently.
World models are essential for AI to be able to be trusted. They allow us to understand whether the AI’s success or failure was due to the right reasons, and not simply because there was a misalignment between us as humans giving it instructions and its understanding of the environment it was operating in.
Relationship between agents and LLMs
We previously said that agents do not need to have AI to meet the above definition of an agent. This can be extrapolated further to say that AI agents do not need to have an LLM core. Agentic and multi-agent systems may not include any Gen AI at all. This is easy to see using our previous definitions of autonomy and agency: clearly LLMs are not required to enable either of these concepts.
It might seem obvious to say agents don’t have to be LLMs, but most of the examples of agents that people mention today do have an LLM core. It is also worth mentioning that often things with an LLM core are called agents, but do not have the ability to exercise agency at all. As a result, these would not meet our criteria for an agent.
Agents and multi-agent systems have been a cornerstone for AI for well over 30 years. The reason why agentic architectures have taken off in the LLM era is because LLMs provide a rich and natural way for humans to communicate goals with AI systems, and natural language provides a way for agents to communicate with each other. The classic phrase of human language being the way to communicate about anything in an inefficient and imperfect way, rings true here. , rings true here.
Five additional dimensions of multi-agent systems (MAS)
We can now look a little deeper into what a multi-agent system is and how we can classify it. On one hand, we can describe agents and their properties of autonomy, agency, and authority. On the other, we can describe dimensions of the whole system.
Here, we propose five dimensions that help us better understand multi-agent systems. The first dimension is size. Then, we talk about heterogeneity. While homogeneous systems are those where agents share similar roles (often called swarms), heterogeneous systems feature specialized agents that handle complex tasks. Heterogeneous systems can self-organize and coordinate to solve a problem but require sophisticated coordination. We then consider the concept of centralization. Centralized systems require rigid structures and orchestration, but are more controllable and explainable. Decentralized systems distribute decision-making broadly, enhancing scalability and resilience, but complicating system coherence and control. These three dimensions may seem like the more the merrier, but larger, heterogeneous, and decentralized systems are harder to control.
Now let´s go back a little to describing aspect of agents rather than the system with specialization. Generic agents often exhibit greater autonomy, capable of flexible decision-making in diverse scenarios, but are rarely able to complete complex tasks. Specialized agents, whilst highly skilled in specific domains, typically exhibit higher agency and lower authority, executing only narrowly defined tasks. The reason why a dimension at an agent level is in this section is because while agentic systems often have a mix of specialization of agents, the system itself can also be described in terms of specialization too.
Lastly, there is the degree to which the system is deterministic or not. Determinism describes how rigid and predictable a system is. Basically, if you do the same thing multiple times, a deterministic system provides the same answer every time. This is where we are seeing lots of change with the wave of Gen AI. Typically, agentic AI systems have been very deterministic. If the thermometer detects a level at 20°C, then it will turn the heating off. Therefore, a fully deterministic system will always produce the same outputs given the same inputs. Their performance will always be the same, which is both good and bad. By contrast, non-deterministic systems might adapt and change their behavior over time. This allows them to improve over time, but also runs the risk that their behavior might become worse or even harmful. It is therefore important to understand how to manage this emergent behavior and monitor it to ensure the desired emergent behavior is obtained.
These dimensions interact in intricate ways, and understanding them is key to designing multi-agent systems that have the desired performance and trustworthiness across diverse architectures and use cases.
To learn more about these topics and explore them further, visit Robert Engels’ blog here.
Maturity model for autonomous AI enablement
We have spoken a lot about what agents are and what an agentic AI system is, but how can we understand them better? Understanding the degree of agency and autonomy in a system is vital to understanding both its power and its risk profile. For example, if we take well-known agents that are used by humans today such as real estate agents, travel agents, and insurance agents, we can plot how much autonomy and agency we give them. We can also understand why we would not want to give full autonomy or agency to them.
We want a real estate agent to be autonomous in selling our homes, but to not have the agency to agree to the sale price without us. We might give travel agents agency within relatively tight constraints to make bookings on our behalf, but not make major changes to dates or destinations. We would want insurance agents to have reasonably high autonomy and agency; they can take out insurance, make sure we have the coverage we need, and we trust they are more competent at that than us.
If we look in the extremes, we find high agency with low autonomy such as sports agents. They negotiate contracts and agree terms, but only when the athlete gives them permission to talk to someone. An extreme example of high agency with high autonomy is a secret agent. Here, the mission is provided, but the agent can decide entirely how they complete it and have full agency, even beyond the law, to act however they choose to achieve the outcome. Hopefully this section helps you realize that the level of autonomy and agency we give to human agents in our world today is the result of the decisions we make and the desired outcomes we have in mind. We must think of autonomous agentic AI systems with the same clarity. These systems will perform within the bounds we give them and optimize against the purpose we assign to them.
A complex landscape
Whilst the mainstream narrative talks about agent implementations as a simple architectural pattern, our exploration of the many attributes and dimensions of agents shows that this is a much deeper topic. Autonomous systems and AI agents will be a defining feature of the technological landscape of our future, and understanding the qualities and dimensions of agency will help us navigate this complex and exciting future with confidence
About the Capgemini Group AI Lab
We are the AI Futures Lab, expert partners that help you confidently visualize and pursue a better, sustainable, and trusted AI-enabled future. We do so by understanding, pre-empting, and harnessing emerging trends and technologies to ultimately make trustworthy and reliable AI that triggers your imagination, enhances your productivity, and increases your efficiency. We will support you with the business challenges you’re currently aware of and the emerging ones you will need to be aware of to succeed in the future. We create blogs, like this one, Points of View (POVs), and demos around these focus areas to start a conversation about how AI will impact us in the future. For more information on the AI Lab and more of the work we have done, visit this page: AI Lab.
