The New Industrial Revolution: the application of Autonomous Artificial Intelligence (Agentic Artificial Intelligence) and its collaboration with humans

SumoPPM Agentic AI

Agentic AI, also known as autonomous AI, is radically changing the artificial intelligence landscape by surpassing the capabilities of traditional generative AI. While generative AI is limited to producing content based on pre-established patterns and specific queries, Agentic AI — thanks to its ability to 'reason' based on context — learns as tasks are carried out and acts according to the situation. This enables it to make autonomous decisions and alert the user when it deems it necessary to suggest a change or make a decision regarding an assigned task (Floridi 2025). This is a significant advantage over conventional Generative AI and other technologies, as it enables the automation of repetitive and complex tasks, freeing up human workers' time, boosting their productivity and reducing the need for human supervision.

With this shift in AI capabilities, we are moving from having what we might call 'a single advisor' answering your questions (Generative AI) to a team of experts solving a problem for you or handling a multi-task process (AI Agent systems). Systems built with multiple AI agents in a collaborative framework are useful because they can solve issues involving several tasks from different areas effectively by sharing information and developing the tasks in which each agent specialises to achieve a common outcome (Gridach 2025).

On a technical level, the evolution from generative AI to AI agent systems is also a qualitative leap in computer architecture. While tools such as ChatGPT respond to prompts, AI agents combine contextual 'sensing', predictive analytics, and the repeated execution of actions (Gridach 2025; Mukherjee 2025).

Definition and main features

Agentic AI has the autonomy to achieve long-term goals by organising multitasking processes that adapt to unforeseen conditions through a combination of environmental sensing and contextualised reasoning, devising novel solutions as it does so. This represents a significant change in how digital systems interact with society (Mukherjee 2025). Unlike traditional generative models, these autonomous agent systems possess four key attributes:

1. Goal-based reasoning: LLMs (Large Language Models) are used as the backbone of the system to facilitate the segmentation of complex goals into smaller ones that are assigned to different agents (Ghose 2024).

2. Trial-and-error learning: This involves real-time feedback mechanisms that enable strategies to be adapted in response to environmental changes. These mechanisms employ techniques such as metacognition[1] nd Bayesian optimisation[2], as well as techniques such as Reinforcement Learning from Human Feedback (RLHF). The latter involves the system learning from the feedback given by the human user when using and/or monitoring the system (Gridach 2025; Mukherjee 2025).

3. Multi-agent cooperation: a number of specialist agents are employed to manage each task in the sequence of steps, replicating human organisational frameworks, enabling collaborative problem solving or multi-step processes (Ghose 2024; Gridach 2025).

4. Metacognition[3]: The ability of an AI agent system to monitor and critically evaluate its own performance through algorithmic self-reflection modules[4] (Miehling 2025).

   
   

The 'degree of intelligence' achieved by these multi-agent systems, or by agents driven by their own Large Language Model (LLM)-based agent models, comes much closer to several characteristics that we consider fundamental to human intelligence when combined, such as autonomy, learning capacity, memory, comprehension, planning and decision-making and execution.

Architectural Framework

The operational structure of these systems is organised in four interconnected modules:

Contextual Perception:

It combines diverse data through specialised application programming interfaces (APIs), merging structured data from IoT sensors and databases with unstructured information inputs such as natural language and satellite imagery (Floridi 2025; PwC 2024). Next-generation transformer models reveal underlying patterns using 'attention mechanisms' (MFAIA 2023; PwC 2024). This technique optimises the ability of artificial intelligence models to focus on relevant information while filtering out contextual distractions.

Dynamic Reasoning:

LLMs (Large Language Models) can develop sequences of thought comprising several steps. For instance, upon detecting a research request, an academic AI agent could generate a hypothesis, identify gaps in the literature on the subject, and plan virtual experiments (Ghose, 2024; Gridach 2025).

Sequences of steps combining deep neural networks[5] and symbolic logic are used to break down complex problems into sequential subtasks that can be solved incrementally to reach the final goal (MFAIA 2023). Frameworks such as Neuro-Symbolic facilitate the integration of deep learning with interpretive capabilities, thereby overcoming the limitations of purely statistical approaches and improving AI decision-making processes (Rodriguez 2021).

Autonomous Execution:

Digital actors, such as software robots and API interfaces, translate algorithmic decisions into concrete actions. Systems such as OpenAI's 'Operator' are demonstrating an emerging capability to manage complex workflows that encompass multiple platforms (Editorial staff of the Luca de Tena Foundation's Journalism Lab 2024).

Continuous Learning:

Learning mechanisms such as Reinforcement Learning from Human Feedback (RLHF) can be employed (Gridach 2025; Mukherjee 2025). In other words, the system learns from human user feedback without requiring continuous human supervision. In the financial sector, for example, Agentic AI systems optimise investment portfolios by evaluating past performance and updated risk preferences.

Transformative applications of Agentic AI across multiple fields

Scientific research:

In academia, multi-agent systems are used to collaborate with human researchers and enhance their capabilities (Gridach 2025).

• One useful application is automated literature review, although there is room for improvements in this area. This is essential for summarising literature and identifying research trends in a given scientific field.

• They can also be used to propose novel hypotheses using generative concept maps[6].

• They can also be used to identify knowledge/research gaps in a given field.

  
  

Optimisation of educational services:

University institutions are implementing multi-agent ecosystems to:

• Personalise educational routes through predictive analysis of competences.

• Automate complex administrative processes (e. g. enrolments, validations).

• Generate adaptive pedagogical content for users using a conversational approach.

  
  

Transforming customer service:

AI Agent systems are reshaping commercial relationships by:

• Autonomous resolution of a large proportion of incidents, eliminating the need for human intervention.

• Anticipating needs by analysing customer sentiment in real time.

• Unified, multi-channel integration (e.g. email, social media and CRM) (PwC 2024).

  
  

  Companies can reduce operational costs by using specialised micro-agents to answer frequently asked questions, process returns, build customer loyalty and more.

  
  

Optimising logistics and inventory management:

AI systems in logistics can adjust delivery routes in real time in the event of problems such as traffic or adverse weather conditions, ensuring fast deliveries and reducing costs. AI systems can also predict changes in demand and automate inventory ordering, avoiding stock-outs or overstocking by analysing purchasing trends and seasonal events.

Meeting coordination and agenda management:

AI agents improve meeting organisation by coordinating agendas, booking convenient times and sending reminders. During meetings, they transcribe conversations, highlight key points and generate summaries. They also assign tasks agreed upon during the meeting, monitor their progress and report any delays, thereby improving efficiency, particularly within international teams.

Personalising retail and entertainment experiences:

In the retail sector, Agentic AI systems analyse customer data, such as preferences and shopping behaviour, in order to provide personalised recommendations and marketing campaigns. Companies such as Coca-Cola and Netflix use this technology to tailor their products and content to consumers' needs. In physical shops, AI uses augmented reality to guide shoppers through apps on their mobile devices, displaying location-based offers and product details.

These are just a few examples of the many applications of AI agent systems. Their versatility allows them to be integrated into diverse contexts such as healthcare, finance, logistics and entertainment. By automating everyday tasks, enhancing critical processes and providing actionable insights, these systems increase operational efficiency, enable new business models and improve the end-user experience.

Ethical challenges and the emerging regulatory framework

Dilemmas of attribution of responsibility:

As a constantly developing technology, autonomous AI still requires specific regulatory frameworks. The operational autonomy of Agentic AI gives rise to ethical issues concerning legal liability and algorithmic transparency, as responsibility becomes diluted among developers, end users, and the systems themselves (Floridi 2025).

The EU's Artificial Intelligence Act 2024 requires high-risk systems, such as those used in banking or medicine, to include XAI (Explainable Artificial Intelligence) mechanisms and human oversight (Alexander Thamm 2025). Technologies such as digital twins enable continuous auditing by replicating operational scenarios to identify potential biases or failures. Digital twins in AI are virtual replicas of real-world objects, systems, or processes that use real-time data, machine learning, and reasoning to accurately mirror and analyse their physical counterparts. They constantly collect, process and store detailed operational data, providing a transparent and traceable record of system behaviour and changes over time.

However, there are still challenges in areas such as intellectual property. When an AI agent generates patents or creative content, questions of ownership arise. At the European level, ownership is generally considered to belong to the user of the AI tool (European Innovation Council and SMEs Executive Agency 2024).

Responsible implementation strategies:

The recent AAIO (Agentic AI Optimisation) framework highlights the need to develop the following (Floridi 2025):

• Standards for autonomous AI governance; ensuring interoperability between digital platforms.

• Mechanisms for oversight and transparency of digital interactions must be put in place.

• It also calls for algorithm systems that can be publicly audited, with established parameters for decision-making, to bring reliability to users (Floridi 2025; Mukherjee 2025).

  
  

  Interdisciplinary collaboration between engineers, programmers, philosophers and policymakers is emerging as a prerequisite for navigating this new technological frontier (Floridi 2025).

  
  

Conclusion

The evolution of agentic AI is radically redefining automation paradigms, presenting both unprecedented opportunities and complex challenges. Although technical advances in autonomous reasoning and multi-agent collaboration have the potential to transform key sectors, the academic community and its developers must adopt a critical and proactive stance on the ethical governance of these systems. The development of emerging theoretical frameworks and regulatory proposals is an initial step towards the responsible use of this transformative technology.

However, it should be noted that the effectiveness of an AI agent system (Agentic AI) depends on several factors, including the quality of the data used, the system's architecture, how digital content and information are structured, and the mutual reinforcement provided by the human user through continuous use and monitoring of the data and results to fine-tune quality. In other words, human users and programmers still play a role in this production and knowledge chain, which cannot be ignored. The efficiency of the agents also depends on the instructions provided, the relevance of the information, and the quality of the data supplied, all of which is the responsibility of human users and programmers. This approach will help to ensure that AI-human collaboration achieves good results.

Notes

[1] Metacognition in AI is the ability of AI systems to ‘reflect on, monitor and adjust their learning models’ (Wei 2024).

[2] Bayesian optimisation is a technique that helps to find the best solution to a problem without the need to test every option. It is particularly useful for tuning the parameters of artificial intelligence models, thereby saving time and resources (Amat Rodrigo 2020).

[3] See footnote 1.

[4] These are pieces of software that allow an AI system to analyse and adjust its own processes automatically.

[5] Deep neural networks are systems that mimic the way the human brain works. They use many connected layers to process information step by step (UNIE Universidad 2024).

[6] These are visual representations of concepts and their interrelationships, produced by artificial intelligence.

Bibliography

Alexander Thamm. 2025. "The State of AI 2025. Agentic AI and Beyond." alexanderthamm.com. Accessed May 2025. https://www.alexanderthamm.com/en/blog/the-state-of-ai-2025/.

Amat Rodrigo, Joaquín. 2020. "Optimización bayesiana de hiperparámetros." CIENCIADEDATOS.NET. April. Accessed May 22, 2025. https://cienciadedatos.net/documentos/62_optimizacion_bayesiana_hiperparametros#:~:text=Con%20esta%20estrategia%2C%20se%20consigue,eligiendo%20%C3%BAnicamente%20los%20mejores%20candidatos.

European Innovation Council and SMEs Executive Agency. 2024. "European Comission." European Comission Intellectual Property Helpdesk . July 2024. Accessed May 25, 2025. https://intellectual-property-helpdesk.ec.europa.eu/news-events/news/artificial-intelligence-and-copyright-use-generative-ai-tools-develop-new-content-2024-07-16-0_en.

Floridi, Luciano, Buttaboni, Carlotta, Hine, Emmie, Morley, Jessica, Novelli, Claudio & Schroder, Tyler. 2025. "Agentic AI Optimisation (AAIO): what it is, how it works, why it matters, and how to deal with it." Arxiv 1-15. https://arxiv.org/abs/2504.12482.

Ghose, Shomit. 2024. "The Next “Next Big Thing”: Agentic AI’s Opportunities and Risks." UC Berkeley Stuardja Center for Entrepeneurship & Technology. Edited by Berkeley. December 19. Accessed May 22, 2025. https://scet.berkeley.edu/the-next-next-big-thing-agentic-ais-opportunities-and-risks/.

Gridach, Mourad, Nanavati, Jay, Zine El Abidine, Khaldoun, Mendes, Lenon & Mack, Christina. 2025. "Agentic AI for Scientific Discovery: A Survey of Progress, Challenges, and Future Directions." Arxiv 1-14. chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://arxiv.org/pdf/2503.08979 .

MFAIA. 2023. "Universidad Politécnica de Madrid." Máster en fundamentos y aplicaciones de la Inteligencia Artificial. May 13. Accessed May 27, 2025. https://mfaia.dia.fi.upm.es/aprendizaje-automatico-la-magia-detras-de-la-inteligencia-artificial/.

Miehling, Erik, Ramamurthy, Karthikeyan, Varshney, Karthikeyan, Riemer, Matthew, Bouneffouf, Djallel, Richards, John T., Dhurandhar, Amit, Daly, Elizabeth, Hind, Michael, Sattigeri, Prasanna, Dennis, Ambrish Rawat, Wei, Gajcin, Jasmina & Geyer, Werner. 2025. "Agentic AI Needs a Systems Theory." Edited by IBM Research. ARXIV.org 1-15. Accessed May 2025. chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://arxiv.org/pdf/2503.00237.

Mukherjee, Anirban and Chang, Hannah. 2025. "Agentic AI: Autonomy, Accountability, and the Algorithmic Society." SSRN 1-28. Accessed May 2025. https://ssrn.com/abstract=5123621 or http://dx.doi.org/10.2139/ssrn.5123621.

PwC . 2024. "Agentic AI – the new frontier in GenAI." PwC . October 27. Accessed May 2025. https://www.pwc.com/m1/en/publications/agentic-ai-the-new-frontier-in-genai.html.

Editorial staff of the Luca de Tena Foundation's Journalism Lab. 2024. "Laboratorio de periodismo Fundación Luca de Tena." laboratoriodeperiodismo.org. November 19. Accessed May 27, 2025. https://laboratoriodeperiodismo.org/openai-anuncia-operator-el-asistente-de-inteligencia-artificial-que-busca-transformar-la-interaccion-tecnologica/.

Rodríguez, Samuel. 2021. "Big Data Magazine." November 19. Accessed May 27, 2025. https://bigdatamagazine.es/que-es-la-inteligencia-artificial-neurosimbolica/.

UNIE Universidad. 2024. "UNIE Universidad." Universidad UNIE. julio 9. Accessed mayo 2025. https://www.universidadunie.com/blog/que-son-redes-neuronales.

Wei, Hua, Shakarian, Paulo, Lebiere, Christian, Draper, Bruce, Krishnaswamy, Nikhil & Nirenburg, Sergei. 2024. "Metacognitive AI: Framework and the Case for a Neurosymbolic Approach." arXiv s. p. Accessed May 22, 2025. https://arxiv.org/abs/2406.12147.