By
Javier Surasky
In the current debate on AI, public attention has focused on generative models and the consequences of their capabilities to produce text, images, video, or code. Behind that focus, however, a form of AI is emerging that responds to human requests and can act autonomously, carrying out tasks and making chained decisions in complex environments: agentic AI (WEF, 2025; IBM, n.d.).
Currently in advanced experimentation and early
deployment in areas such as organizational management, digital logistics,
cybersecurity, and the provision of public and private services (BCG, 2025;
McKinsey & Company, 2025; WEF, 2025), agentic AI entails both an increase
in technical capabilities and a qualitative shift in how decision-making is
delegated, with direct implications for politics, power asymmetries, and the
possible frameworks for global AI governance (HLAB-AI, 2024; OECD, 2023).
Nevertheless, what is agentic AI? From a conceptual
standpoint, these are action-oriented AI systems, or systems with agency (hence
the name), that depart from traditional models that generate outputs in
response to discrete instructions to become actors in the real world. “An AI
agent is a system that uses AI and tools to carry out actions in order to
achieve a given goal autonomously” (Bornet et al., 2025:15). As such, it can
define goals, plan sequences of actions, interact with digital tools, and evaluate
outcomes in order to adjust its behavior (Russell & Norvig, 2004; Sapkota
et al., 2025) without the need for human interaction or guidance.
The reference to the agent, defined as one who has the
capacity to act, is far from an anthropomorphic metaphor. Instead, it describes
the system’s distinctive functional property in relation to its predecessors.
Its theoretical foundations lie in the convergence between research on agents
and multi-agent systems, foundation models, and large language models (McShane
et al., 2024; Labaschin, 2023). These operate within layered architectures that
integrate components for reasoning, organization, and action execution,
connected through standardized protocols to data sources or other agents (WEF,
2025; OpenAI, n.d.), resulting in a distributed decision-making system capable
of operating continuously within changing digital environments.
Agentic AI thus represents both a rupture and a
continuity. Expert systems, algorithmic trading platforms, or automated
industrial infrastructures have long incorporated some degree of decision
delegation, but agentic AI generalizes, integrates across sectors, and scales
this capability.
For these reasons, rather than an absolute historical
novelty, agentic AI should be understood as a systemic advance that produces a
qualitative shift in AI models while also redefining their very nature.
What we observe is a deepening of the politics of
automation, oriented toward goals such as efficiency, effectiveness, speed of
response, and control over deployed actions, all within a framework of
increasing systemic density in decision-making processes and their
implementation.
At this point, and before examining the potential
impact of advanced agentic AI, it is necessary to pause and reflect on the
central idea of “delegation of decisions.”
In general terms, one can distinguish operational
delegation—understood as the automated execution of tasks—from tactical
delegation, which involves optimization, coordination, and the selection of
courses of action within predefined objectives, and strategic delegation, where
what is transferred is the capacity to define objectives and priorities.
Agentic AI currently operates mainly within tactical delegation, where systems
acquire the capacity to coordinate complex processes and make chained decisions,
without this necessarily implying a transfer of strategic control. In other
words, these systems enjoy functional autonomy but not political or normative
autonomy, as they operate under external organizational incentives.
Even with these limitations, the advent of agentic AI
introduces inevitable “ruptures” or discontinuities that warrant attention. We
move from working with singular models to operating within ecosystems of
agents, where the ability to coordinate and work jointly is as critical to the
model as individual performance. Moreover, collaborative capacity becomes part
of individual performance evaluation: a system that produces good results in
isolation but not when interoperating with others will be deficient from the
standpoint of agentic AI (WEF, 2025; Schick et al., 2023).
The emphasis shifts from content production toward
direct participation in processes, ranging from resource allocation and
workflow management to team-based task execution and integration with
pre-existing systems (AWS, n.d.; IBM, n.d.). This shift displaces humans from
action executors to supervisors, auditors, and controllers of AI agents,
requiring a redesign of organizational processes and the development of new
regulatory frameworks (OECD, 2023; HLAB-AI, 2024).
To make this more concrete, consider an example:
according to recent reports published by companies, consultancies, and research
centers (WEF, 2025; McKinsey & Company, 2025; OpenAI, n.d.), agentic AI is
already coordinating specialized agents in activities such as inventory
management, logistics, customer service, or the maintenance of digital
infrastructures. These systems respond to real-time needs, directly executing
actions such as reallocating resources or reshaping workflows in response to changes
in external conditions. This deployment of tactical response capabilities at
ever-increasing speed allows organizations that adopt agentic AI to achieve
greater efficiency in their regular operations.
The combination of relative (tactical and functional)
autonomy, increased execution speed, and the inherent scalability of agentic AI
is already reshaping decision-making logic, especially in sectors where rapid
and coordinated responses are critical.
However, this shift generates political and security
consequences that go beyond the purely technical (Mitre & Predd, 2025).
Here, the political significance of agentic AI becomes clearer: by expanding
and strengthening capabilities already considered strategic in the digital
economy, security, or public policy management, among other sectors, the new
agent role assumed by AI produces competitive advantages in highly complex
environments, at the cost of reinforcing dependence on digital infrastructures
and timely access to data (UNCTAD, 2024; ITU, 2025).
That is not a process articulated exclusively through
states, but one that depends on public–private ecosystems, often transnational
in nature and heavily weighted toward the corporate sector, which enjoys the
“competitive advantage” of operating without the constraints that borders
impose on states. As a result, the power associated with agentic AI exceeds the
state’s capacity for action and rests with those who exercise effective control
over the decision architectures that shape its operation. Put more simply,
power lies with those who design how the system makes decisions, not with those
who merely benefit from them.
Once this element is made explicit, it becomes evident
that the development of agentic AI reinforces pre-existing structural
asymmetries (UNCTAD, 2024). For many technologically lagging countries—or even
those with relatively high levels of development but operating outside the
technological frontier—the adoption of agentic AI may entail increasing
dependence on externally defined decision architectures, with minimal margins
for adaptation, appropriation, and sovereign control (Srivastava & Bullock,
2024; Colomina Saló & Galceran-Vercher, 2024). The “coloniality of being”
thus becomes embodied.
It is worth pausing on this idea, which once again
highlights the need to think about AI through the lens of the social sciences.
The concept of the coloniality of being is one of the pillars of decolonial
thought. It is defined as “the radical betrayal of the trans-ontological
through the formation of a world in which the non-ethics of war are naturalized
through the idea of race” (Maldonado-Torres, 2007:267).
The trans-ontological presents a primordial ethical
relation in coloniality, namely the gift bestowed by the colonizing Self upon
the colonized Other. This relation of imposed superiority of the “Self” through
the degradation of the “Other” is expressed as a foundational betrayal of
coloniality. In the coloniality of being, it manifests through the colonized
subject’s acceptance of the naturalness of this logic, resulting in an order in
which “the non-ethics of war”—that is, murder, rape, and so forth—are justified
by the concept of “race” (with the colonizer’s race deemed superior and the
colonized’s inferior). It should be specified, however, that race in this
quotation serves as a starting point for the inclusion of other variables, such
as gender. What we argue, then, is that agentic AI could give physical form to
a relationship that, until now, had been visible only in its consequences, yet
remained incorporeal.
This idea can be reinforced by considering the
adoption of standardized agentic architectures in cultural contexts marked by
different worldviews, values, and unequal institutional capacities. The
integration of agentic AI systems into critical digital infrastructures—such as
energy supply control platforms or economic coordination systems—leaves states
lacking the capacity to design their own decision architectures at the mercy of
those established by dominant actors, reproducing the “Self–Other” relationship
in digital terms. Under ostensibly well-founded pretexts of interoperability,
such systems end up generating subordination to externally imposed decision
schemes, conditioning domestic tactical decisions on design logics,
optimization criteria, and normative frameworks embedded in agentic systems,
thereby dealing a new blow to digital justice and sovereignty (UNCTAD, 2024;
ITU, 2025). In this sense, the risk associated with agentic AI is structural
and probabilistic rather than deterministic, and depends on specific social,
cultural, political, and institutional configurations.
As a result, it is possible to speak of a “geopolitics
of the delegation of decisions” as an analytical category, but not as a theory
or deterministic prediction. Accordingly, we do not claim that agentic AI has
already transformed the international order; rather, we limit ourselves to
noting that it introduces a plausible trajectory for the reconfiguration of
power associated with the delegation of tactical decisions. That is reflected
in a shift in how power is exercised in contexts where such decisions are
particularly consequential, and in the definition of decision architectures as
a resource of international projection with, now truly, strategic value.
The changes generated by this geopolitics of decision
delegation appear at multiple levels. Economically, it grants a competitive
advantage to those who can automate and coordinate complex processes more
efficiently (McKinsey & Company, 2025; BCG, 2025). In the security domain,
it creates opportunities for automating sensitive functions. More broadly
speaking, it gives rise to risks of systemic errors, especially when high
levels of interaction exist among actors with conflicting objectives in contexts
of high uncertainty (Mitre & Predd, 2025).
Nevertheless, not everything associated with agentic
AI entails risks and pressures in international geopolitics. With appropriate
management, governance, and expert support, it is possible to advance global
democratic processes of technical standardization, define equitable
interoperability frameworks, and even reduce barriers to access to cutting-edge
technologies. None of this depends on the technology itself, but rather on
political decisions and on the capacity of actors with differing interests to
participate in and influence deliberative processes.
Here we return to an issue we will merely mention, as
it has been addressed in previous blog posts: the early deployment of agentic
systems allows first movers to set de facto standards that subsequently
constrain others, producing the legal phenomenon of “regulatory capture.” This
concept has been known for decades and is defined by Carpenter and Moss
(2014:13) as “the result or process by which regulation, in law or application,
is consistently or repeatedly directed away from the public interest and toward
the interests of the regulated industry, by the intent and action of the
industry itself.”
However, if progress in AI governance more broadly is
already slow and challenging due to conflicts among major powers—creating a
significant gap between global normative-institutional development and
technological advance—this problem is exacerbated in the specific domain of
agentic AI. Most existing regulatory instruments focus on individual models or
applications and are ill-suited to distributed, adaptive systems that interact
continuously with other systems and agents (HLAB-AI, 2024; OECD, 2023).
That is compounded by the fact that “current
regulations are unable to cope with the AI revolution due to the pacing problem
and the Collingridge dilemma” (Tehrani, 2022:21). This dilemma expresses an
inherent tension in the field of technology: in the early phases of a new
technology’s development, it is relatively easy to modify or regulate it, but
difficult to foresee its social, economic, or political impacts; in its mature
phase—when those impacts become visible—the technology is already embedded in
global social and economic life, making regulation extremely difficult or
politically unviable (Collingridge, 1980:17–18).
For these reasons, agentic AI is not merely “another
advance” in the field of AI, but a qualitative change that enables new forms of
exercising power through digital systems. As we have noted, it is no longer
about controlling others’ decisions, but about taking control over how
decisions are made.
Those who fail to understand and participate in the
governance of agentic AI at the international level will have to accept the
risk of formulating their policies through someone else’s lenses—without even
realizing it.
References
AWS (Amazon Web Services). (n.d.). What is Agentic
AI? https://aws.amazon.com/what-is/agentic-ai/
BCG (Boston Consulting Group). (2025). AI Agents.
https://www.bcg.com/capabilities/artificial-intelligence/ai-agents
Bornet, P., Wirtz, J., Davenport, T. H., De Cremer,
D., Evergreen, B., Fersht, P., Gohel, R., & Khiyara, S. (2025). Agentic
Artificial Intelligence: Harnessing AI Agents to Reinvent Business, Work, and
Life. World Scientific Publishing.
Carpenter, D., & Moss, D. A. (Eds.). (2014). Preventing
Regulatory Capture: Special Interest Influence and How to Limit It.
Cambridge University Press.
Collingridge, D. (1980). The Social Control of
Technology. Frances Pinter.
Colomina Saló, C., & Galceran-Vercher, M. (2024).
The other geopolitics of artificial intelligence. CIDOB Journal of
International Affairs, (138), 27–50.
HLAB-AI (UN High-Level Advisory Body on Artificial
Intelligence). (2024). Governing AI for Humanity. United Nations. https://www.un.org/sites/un2.un.org/files/governing_ai_for_humanity_final_report_en.pdf
IBM. (n.d.). What is Agentic AI? https://www.ibm.com/think/topics/agentic-ai
ITU (International Telecommunication Union). (2025). AI
Standards for Global Impact: From Governance to Action. https://www.aigl.blog/ai-standards-for-global-impact-itu-2025/
Labaschin, B. (2023). What Are AI Agents? When and
How to Use LLM Agents. O’Reilly Media.
McKinsey & Company. (2025). Seizing the Agentic
AI Advantage. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/seizing%20the%20agentic%20ai%20advantage/seizing-the-agentic-ai-advantage.pdf
McShane, M., Nirenburg, S., & English, J. (2024). Agents
in the Long Game of AI: Computational Cognitive Modeling for Trustworthy,
Hybrid AI. MIT Press.
Mitre, J., & Predd, J. B. (2025, February 10). Artificial
General Intelligence’s Five Hard National Security Problems. RAND
Corporation. https://www.rand.org/pubs/perspectives/PEA3691-4.html
OECD (Organisation for Economic Co-operation and
Development). (2023). Advancing Accountability in Artificial Intelligence.
OECD Digital Economy Papers, (349). https://www.oecd.org/content/dam/oecd/en/publications/reports/2023/02/advancing-accountability-in-ai_753bf8c8/2448f04b-en.pdf
OpenAI. (n.d.). A Practical Guide to Building
Agents. https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
Russell, S. J., & Norvig, P. (2004). Artificial
Intelligence: A Modern Approach (2nd ed.). Pearson.
Sapkota, R., Roumeliotis, K., & Karkee, M. (2025).
AI agents vs. agentic AI: A conceptual taxonomy, applications, and challenges. Information
Fusion, 126.
Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R.,
Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., & Scialom, T.
(2023). Toolformer: Language models can teach themselves to use tools. Proceedings
of the 37th Conference on Neural Information Processing Systems (NeurIPS). https://proceedings.neurips.cc/paper_files/paper/2023/file/d842425e4bf79ba039352da0f658a906-Paper-Conference.pdf
Srivastava, S., & Bullock, J. (2024). AI,
Global Governance, and Digital Sovereignty. https://arxiv.org/pdf/2410.17481
Tehrani, P. M. (Ed.). (2022). Regulatory Aspects of
Artificial Intelligence on Blockchain. IGI Global.
UNCTAD (United Nations Conference on Trade and
Development). (2024). Digital Economy Report 2024. https://unctad.org/system/files/official-document/der2024_en.pdf
WEF (World Economic Forum). (2025). AI Agents in
Action: Foundations for Evaluation and Governance. https://reports.weforum.org/docs/WEF_AI_Agents_in_Action_Foundations_for_Evaluation_and_Governance_2025.pdf