How do we measure intelligence? How do we determine that one individual or program is more intelligent than another? Is it based on the amount of information they possess? Or perhaps on how they use that information to find the best solutions? If you think it’s the latter, you likely believe that the most intelligent entity is the one with the greatest autonomy—and that has everything to do with agent-based artificial intelligence.

AI was inspired by neural networks, morphologically speaking: through layers of the brain of varying depths, neurons connect in networks and, in doing so, store information in memory and, consequently, learn. This is, so to speak, the physiological aspect of intelligence.

However, intelligence is not measured solely by the ability to quickly form many connections and store large amounts of data. It is, to a large extent, measured by the decisions we make—more specifically, by the quality of the choices we make based on the information we have stored. To make better or worse decisions, we need a certain degree of autonomy; or, to use another term, agency.

When we introduced artificial intelligence agents here on the Inmetrics blog, we described their key characteristics: “social” ability, reactivity, proactivity, knowledge base, intention, innocence, and autonomy. These attributes define how agentic an artificial intelligence is. In other words, “agent-like” is an adjective that describes an AI and varies in intensity: some artificial intelligence agents are more agent-like than others.

Explaining the concept

Perhaps the best way to explain what agent-based AI is would be to think about the interactions we often have with chatbot-style AI agents. You ask a question, and it generates a response. You ask another question, and it generates another response, and so on. In other words, these AI applications are designed to solve specific tasks.

An agent-based artificial intelligence is oriented toward one or more objectives. Rather than seeking to perform isolated tasks, they have the ability to plan each step toward the ultimate goal. If we revisit the characteristics of artificial intelligence agents mentioned earlier, agent-based AIs have greater autonomy, proactivity, and intent. In short, they have more agency over the tasks assigned to them.

We can consider applications of agent-based artificial intelligence to be advanced versions of AI systems. Therefore, the conceptual foundations that guided their development were present from the very beginning of discussions on the topic, in the 1940s. However, it was not until 2023, with Andrew Ng’s work on structuring what he called “agent-based workflows,” that agent-based AIs ceased to be mere proposals in scientific articles and gained tangible form in widely used applications.

In his work, Andrew Ng presents four design patterns for agent-based artificial intelligence applications:

1. Reflection → LLMs review their own work, identify errors, and find ways to correct them
2. Use of tools → : LLMs are given capabilities—such as searching the Internet, executing code, or performing other functions—that help them find information, make decisions, or process data
3. Planning → LLMs identify the need and develop a multi-step plan to achieve a goal.
4. Multi-agent collaboration → artificial intelligence agents work together: they discuss options and divide tasks to find a solution that is better than what a single agent could achieve.

Agent-based artificial intelligence: how it works

We present a conceptual overview of agent-based AI applications. From a slightly more technical perspective, these applications are defined by their components, architecture, and workflows.

Among these components, perhaps the key element that distinguishes traditional artificial intelligence agents from agent-based AIs is persistent memory. In addition to using short-term memory—specifically, the context window of LLMs—they maintain long-term memories in vector databases. It is persistent memory that enables the configuration of asynchronous and autonomous routines without the need for an immediate prompt. Thus, agentic AI can operate continuously.

Another key component is the toolbox. There are agent-based AI applications that have the autonomy to create their own virtual machine. By operating it, they can compile and execute code, manipulate files, and access the Internet to browse, search for information, and install dependencies on systems.

From an architectural standpoint, we can say, in simple terms, that agent-based artificial intelligence uses an application architecture that includes various types of artificial neural networks: Transformers as the core processing unit and other peripheral networks, which are generally subtypes of feedforward neural networks, such as convolutional networks. These operate in computer vision modules and are used, for example, when the application needs to read elements on the screen—buttons, images, among others.

As we have already mentioned in this text, it was the structuring of “agent-based workflows” that made the development of agent-based AIs possible. These workflows are their most distinctive feature, determining how these applications function. Broadly speaking, we can divide the workflow into four stages:

1. Perception phase → identifying information—whether from the prompt or other sources—that will be necessary to understand the objective
2. Planning and reasoning phase → the agent sets the goal, breaks it down into sequential tasks, and identifies the tools that will be used to complete each one
3. Action phase → the system executes the actions, orchestrated by the core, typically artificial intelligence agents structured as Transformer-type artificial neural networks
4. Learning phase → The application assesses whether the task deliverables meet the established objective by comparing what has been completed with the desired state. If the deliverables are evaluated negatively, the application seeks a way to correct the error.

Agent-based artificial intelligence is already being used in a wide range of applications, both for corporate and personal use. It is employed for countless activities, ranging from in-depth data analysis to fraud detection in the financial system, as well as managing a computer and all user tasks end-to-end.

Many of these capabilities of agent-based AI are applied in the projects we develop for our clients, led by our Digital Acceleration, Digital Experience, and Consulting, Data, and Artificial Intelligence units. Depending on your needs, we have a specialized unit to assist you, whether in developing your own agent-based AI project—including the testing phase—or in analyzing and implementing artificial intelligence strategies in your systems.

If you’d like to see how agent-based AI can benefit your company’s operations, please contact us! One of our experts will be happy to show you how AI agents have evolved and how your company can take advantage of them!