The hype surrounding generative AI has now been overtaken by agentic AI, which is presented as a response to the limitations of large language models. While generative AI is used to create content, agentic AI is designed to solve complex tasks either autonomously or semi-autonomously. However, the marketing hype often obscures the operational reality. The hype tends to focus the discourse on the promise of novelty, while minimising or even ignoring structural constraints that remain unchanged, as well as the often high development and maintenance costs.
Agentic AI, as IBM indicates, refers to an artificial intelligence system that requires limited human supervision, composed of learning models capable of mimicking certain human decision-making mechanisms to solve problems in real time.
On the other hand, a multi-agent system refers to an architecture in which several agents interact within the same environment. Each agent is designed to perform a specific task, such as searching, analysing or planning, and is coordinated by an orchestration system that organises interactions and ensures the process is consistent overall.
These architectures aim to overcome certain limitations of large language models by fragmenting tasks among several specialised agents, particularly in terms of reliability, consistency, and the reduction of semantic noise, referring to all disturbances in the production, transmission, or interpretation of meaning. One of the most visible and well-documented aspects of this phenomenon is artificial hallucinations, which designate the production of seemingly coherent content that is not based on any verifiable factual reality.
The promises of completely autonomous systems
The promises of agentic AI are part of a broader trend of enthusiasm for systems presented as more autonomous and robust, although this autonomy still largely relies on theoretical assumptions and supervisory conditions rarely emphasised in marketing materials. Indeed, upon closer examination, many of these agents are actually classic automation systems that have been “augmented” by a large language model. Nevertheless, the use of the term “agent” introduces ambiguity about the true agency of these systems. While recent architectures simulate planning stages, their reasoning capacity remains a statistical problem-solving process lacking deep semantic understanding or genuine intentionality.
Most systems described as ‘agent-based’ rely on chains of predefined rules, scripts, and triggers. In these systems, the language model primarily serves as an interface or generation engine. Furthermore, close human supervision is necessary to prevent errors, correct deviations and ensure that operations conform to initial objectives. This significantly limits the autonomy claimed for these architectures. To date, the cumulative operational cost (including inference and human verification) of high-precision tasks can be much higher than that of the human experts these systems aim to support or replace. Additionally, operational responsibility always rests with humans. Therefore, responsibility cannot be transferred in the event of an error; agenticAI remains a tool whose use and effects are under human control.
A continuity, not a break
Multi-agent systems do not depart from the probabilistic foundations of language models, but rather redistribute them across more complex processing chains. The numerous limitations observed in large language models persist. They reconfigure, fragment, and, in some cases, reinforce each other. As with information disorder in the digital age, what deserves scrutiny is not only the technology itself, but also the systemic dynamics it introduces. : cascading propagation, loss of traceability , and illusion of collective validation.
One of the central arguments in favour of multi-agent systems rests on functional specialisation. Each agent is assigned a limited scope of action. Information retrieval, analysis, synthesis, reformulation, and planning are key steps in this process. This approach aims to limit semantic drift by constraining the role of each entity. In practice, this compartmentalisation can indeed reduce certain types of local errors, particularly when agents are strictly governed by business rules or deterministic tools. However, this noise reduction is relative. It depends less on the number of agents than on the degree of external constraint imposed on the reasoning.
When these constraints are lacking, specialisation does not eliminate the uncertainty inherent in probabilistic models. It simply distributes it across several levels of the system.
Increasing the number of agents does not inherently make a system more deterministic, although certain methods of adversarial debate between agents can statistically reduce hallucinations. However, each unit of the system relies on the same probabilistic generation principle: without being anchored in a logical rule engine or a verified knowledge base, increasing the number of actors merely shifts the risk of error to the coordination layer. This dynamic poses a classic problem for complex systems: the cascading propagation of errors.
A snowball effect
In an agentic system, the outputs produced by an agent are rarely treated as hypotheses to be tested. They are most often considered valid inputs for the next agent. This implicit trust forms the basis of a well-documented mechanism in distributed systems engineering.
Agentic AIs are not error-free, and when an error occurs, it rarely remains confined to a single agent. On the contrary, it tends to propagate throughout the processing chain.
Thus, even slightly erroneous information from a search agent can serve as a basis for an analysis agent, which can then be integrated into a final synthesis. At each stage, the error is reformulated, contextualised, and justified. It gains apparent legitimacy, not because it is correct, but because it has passed through several layers of processing. This phenomenon transforms a local failure into a systemic failure, often difficult to diagnose.
Added to this are the problems related to coordinating the agent network. Agentic architectures generally rely on an orchestrating agent that distributes tasks, evaluates intermediate outputs, and decides their sequence. When this orchestrator is itself based on an LLM, it introduces an additional level of uncertainty.
The orchestrator does not assess the accuracy of the outputs it receives. Instead, it assesses their apparent coherence and conformity to an intention or objective expressed in natural language. This can result in coordination biases, feedback loops or the unintentional elimination of critical weak signals.
Therefore, the system’s governance relies on a component that shares the same epistemic boundaries as the agents it supervises. Since exchanges between agents primarily occur through natural-language messages or semi-structured formats, each handover introduces new opportunities for noise, such as ambiguities, divergent interpretations, and context loss.
Unlike a traditional IT pipeline, which is based on strict types and explicit rules, an agentic system relies on successive interpretations. This makes its overall behaviour more difficult to predict, audit, and reproduce.
Towards hybrid systems
Distributed information processing techniques are not new. Computer systems have long relied on modular architectures, in which specialised components perform distinct tasks within structured processing chains. What is changing today is not the principle of functional decomposition, but its application to systems based on large-scale language models.
Therefore, the fundamental problem in agentic AI is not architectural. Rather, it is related to the absence of an internal deterministic foundation. Agents do not manipulate established facts or stable logical rules unless they are explicitly constrained by layers of non-AI code. This is why research and engineering are now moving towards hybrid systems, in which LLMs provide support while validation, decision-making and final structuring rely on deterministic mechanisms.
Translated with the help of Google Translate – humanly post-edited.
To go further
Dwivedi, YK, Helal, MY, Elgendy, IA, Alahmad, R., Walton, P., Suh, A., … & Jeon, I. (2026). Agentic AI Systems: What It Is and Isn’t. Global Business and Organizational Excellence , 45 (3), 253-263. https://doi.org/10.1002/joe.70018
Hosseini, S., & Seilani , H. (2025). The role of agentic AI in shaping a smart future: A systematic review. Array , 26 , 100399. https://doi.org/10.1016/j.array.2025.100399
Kodikara, KASN (2025). Agentic AI Systems: Evolution, Efficiency, and Ethical Implementation. AI Systems Engineering , 1 (2), 23-29. https://doi.org/10.64229/gq9z0p28
Leonardi, P. M. (2025). Homo agenticus in the age of agentic AI: Agency loops, power displacement, and the circulation of responsibility. Information and Organization , 35 (3), 100582. https://doi.org/10.1016/j.infoandorg.2025.100582
Raheem, T., & Hossain, G. (2025, May). Agentic ai systems: Opportunities, challenges, and trustworthiness. In 2025 IEEE International Conference on Electro Information Technology ( eIT ) (pp. 618-624). IEEE. https://doi.org/10.1109/eIT64391.2025.11103638
