Beyond Machine Learning: Is the AI ready to make its own decisions?

In recent years, artificial intelligence or AI has revolutionized the world at all levels, and today we can say that it is part of our vocabulary and routine.

If during 2023 and 2024 we experienced the boom of Generative AI with tools such as Chat-GPT (for text and/or chat generation) or Sora (video generation), 2025 will be the year of Agentic AI (AAI).

What is Agentic Artificial Intelligence?

Relying on Generative AI (GAI) techniques, AI systems are beginning to acquire the ability to create plans and act autonomously in the digital world and are becoming popular as a new way to make Internet applications and services. While GAI systems need a prompt provided by a human – that is, instructions, questions or context information to get the AI system to respond in the direction we want – AAI takes a more autonomous and less user-driven approach, taking process automation to the next level. These systems are capable of executing chained actions – from searching databases or using Internet search engines, to booking online services – and making decisions (including evaluating the veracity of retrieved information, critically analyzing it, or formatting it to meet specific criteria or standards) to achieve goals autonomously and without human supervision.

An AAI system is based on cooperation between so-called agents, which use Generative AI models to operate in unstructured environments, with high uncertainty, and cooperate cohesively to achieve a common goal. They act according to the defined objectives, the cooperation methodology defined by the designer of the application or agentic system, and the degree of freedom they have been given to organize themselves autonomously to optimize the processes.

How does it work?

Intelligent AI agents are still at a very early stage of development, although they are maturing at a rapid pace. That they operate with a greater degree of independence is due to advances in Generative AI, in particular, so-called large language models or LLMs of the GPT-4 type, conversational systems of the ChatGPT type, and techniques for reasoning (such as chain of thought or CoT), information retrieval (such as retrieval with augmented generation or RAG), and knowledge organization (such as knowledge graphs or KG), all of which are based on LLMs.

What capabilities does Agentic AI have because of these advances:

• Persistent memory: Agents can remember previous interactions and use that information to improve their performance, adjusting their responses and decisions based on the user’s history. This capability allows them to maintain consistency in their actions and optimize their behavior over time.

Imagine using an AI assistant to manage your meetings. With persistent memory, the agent remembers your preferred times, regular participants and your preferences about virtual or face-to-face meetings. So, the next time you need to schedule a meeting, the AI will automatically suggest the best options without you having to repeat your preferences each time.

• Autonomous planning: Design and execute long-term strategies based on defined objectives, without continuous prompts. Thanks to advanced decision-making models, they can prioritize tasks and adapt to changes in the environment without requiring direct instructions at every step.

Suppose you want to organize a networking event in your company. An Agentic AI could take care of finding the best available date, reserving spaces, coordinating with suppliers and sending invitations without the need for you to oversee every detail. If an unforeseen event arises, such as the cancellation of a speaker, the AI would autonomously restructure the event to minimize the impact.

• Contextual understanding: They analyze data in real time to make data-driven decisions. This allows them to adapt to dynamic situations, learn from new scenarios and respond more effectively to changes in their environment.

Think of AI-powered customer service. If a user contacts with questions about a product and mentions that he or she has already tried a previous solution, the agent will not repeat the same instructions but will adapt his or her response based on the context. This improves the user experience and optimizes support efficiency.

5 applications and use cases of Agentic AI

• Software development: AI is positioned to transform programming assistants, also called copilots, into smarter software development tools capable of writing larger and more complex code snippets.
• Industry and automation: Autonomous robots powered by intelligent agents will optimize production processes in factories and warehouses, improving operational efficiency and leading to major cost reductions. In intelligent fleet management, agentic AI can make real-time, data-driven decisions to optimize routes and reduce delivery times. It is also a breakthrough in predictive maintenance, as it analyzes large volumes of data to anticipate failures, for example, in sensors, and schedule repairs before critical problems occur.
• Healthcare: AAI is transforming patient care, improving the end-user experience by automating tasks, analyzing medical records and making personalized treatment recommendations. It can also be used in automated diagnostics, interpreting medical images and clinical data to detect pathologies with high accuracy.
• Cybersecurity: AI agents can be used to proactively detect and respond to threats. They monitor networks in real time, detect suspicious patterns and respond to attacks before they can cause damage. Their continuous learning capabilities allow agents to evolve in the face of new cyberattack techniques, adapting more effectively than traditional systems.
• Video games and entertainment: The development of NPCs (non-playable characters) is greatly benefited by the AAI, as they are able to adapt to the player’s behavior and respond based on their actions, offering more realistic and immersive experiences. Interaction becomes much more dynamic.

Challenges we will face in the development of Agentic AI

Despite its enormous potential, the use of Agentic AI (AAI) presents significant challenges that must be addressed to ensure its safe and responsible development. The fact that, in this new paradigm of intelligent software applications, the process or service is provided by a team of intelligent agents cooperating with each other with varying degrees of autonomy, raises the question of what level of autonomy is acceptable from a safety and liability standpoint. Thus, the accelerated growth of its use to create a new layer of applications and services on the Internet raises key questions in terms of governance (including ethics and privacy), reliability and control, which requires adequate oversight and a regulatory framework to enable its responsible integration into society.

1. Agentic AI is expanding without governance and oversight, which can lead to major problems in the long term, such as the generation of undetected biases or the misuse of this technology in critical sectors.
2. As it is still in the early stages of development, AAI makes autonomous decisions, which may not be reliable. These systems may make mistakes, decide on issues for which they have not been validated, make decisions that have not been delegated to them, interpret data incorrectly or fail in unexpected situations. The lack of technological maturity also hinders their integration in sectors that require high levels of accuracy.
3. As IoT advances, so do the threats associated with its malicious use. Cyber-attacks driven by IoT that enable intelligent malware could occur. Unlike traditional malware, these systems could learn from their previous attacks, modify their code to avoid detection, and strategically select targets, exponentially increasing their capacity for damage.

Exploring the future of AI at ARQUIMEA

ARQUIMEA, from its research center located in the Canary Islands, has an Artificial Intelligence Orbital with the aim of responding to the great challenges we face as a society. Some of our research lines are: the acceleration in the search and design of new drugs, the safe autonomy of intelligent agents and robotic platforms and drones, or the implementation of AI systems on board satellites for earth observation and security in space.

In addition, all ARQUIMEA Research Center projects belong to the QCIRCLE project, funded by the European Union and aimed at the creation of a center of scientific excellence in Spain.

“Funded by the European Union. However, the views and opinions expressed are the sole responsibility of the author and do not necessarily reflect those of the European Union and neither the European Union nor the granting authority can be held responsible for them.”