Latest

AI Agents

The society of AI agents: When machines start debating each other

We analyze the society of AI agents and the impact of machines that debate each other. How will employment and science change?

The society of AI agents: When machines debate
The society of AI agents: When machines debate

The society of AI agents is already here. It is a present. But above all it is a present that looks like the future. Better said, assuming the inevitable uncertainty of any future: it looks like different futures, with a strong possibility that in -let's say- ten years reality will be another.

The society of AI agents

But it is worth setting in motion the sociological imagination of what a society organized around the prominence of AI agents will be. An imagination about something that is portrayed as the new technological revolution: mechanization expanded the capacity of humans' physical labor, computers expanded calculation, the internet expanded communication, the society of AI agents expands decision-making capacity.

The society of AI agents is the consequence of the new phase AI has entered. A phase in which, instead of relying on a single model to answer questions or perform tasks in isolation, organizations or networks of autonomous AI systems emerge that collaborate, negotiate and solve complex problems jointly. Each of these organized AI agents can plan, make decisions, use tools and, above all, communicate with other agents and adapt to changing circumstances.

Following the principles of the social division of labor, which the French sociologist Durkheim elevated 123 years ago to the fundamental engine of the social transformation then underway, each agent has a specific function. One can specialize in research; another in coding; another in quality assurance; while another manages workflows or communication with humans. Together they form a dynamic ecosystem in which intelligence is distributed rather than centralized.

The social division of labor

Principles of the social division of labor, for centuries present in organizations made up exclusively of humans, now to be applied to the problems of an increasingly complex society. Goals such as designing a product, managing a supply chain, analyzing financial markets or discovering new drugs require the integration of different areas of knowledge and expertise. A simple AI model can offer some support in some of the phases of the processes aimed at such goals; but a coordinated team of specialized AI agents can address them much more efficiently and, using a fashionable concept, comprehensively.

Let us take a schematic example of a workflow organized around interrelated AI agents. Research agents collect information from various sources, analyst agents evaluate the evidence and identify patterns, planner agents design strategies, engineer agents build solutions, reviewer agents evaluate the quality of such solutions and detect errors, and coordinator agents assign tasks and monitor the progress of the process.

The result is a system capable of handling sophisticated workflows with minimal human intervention, also incorporating flexibility and immediate adaptation to circumstances. In the process, agents exchange information, debate alternatives, verify each other's work and interactively refine the solutions. Instead of relying on a model of “the best answer” (best answer), the system benefits from “discussion” or cross-validation among the different agents. That is, it benefits from the multiple perspectives present, derived from their respective functional specializations.

Scientific activity

The association of AI agents is already the subject of fairly advanced experiments in tool integration (different agents accessing databases, APIs, search engines, external services and software applications), as well as in different industries: software development, healthcare, scientific research, logistics, customer service, financial analysis or law firms.

However, where the greatest impact is expected is in scientific activity, with well-founded hopes for accelerating research on diseases such as Alzheimer's or in quantum matters. A scientific activity where communication between disciplines, the always-desired interdisciplinarity, becomes automatic: a discovery in chemistry, materials science or medicine is immediately reported to biology, renewable energy or pharmaceutical development.

Its experimentation also reveals the problems of this society of AI agents. One of them is that, as the number of agents grows, effective communication and coordination among them becomes more complex, with conflicting decisions emerging. This situation of conflicting decisions can arise when the objectives given to the organization of AI agents overlap with one another and are not hierarchized. After all, they are machines seeking to fulfill objectives. Even when objectives are hierarchized, they can come into conflict. Recall Asimov's three ordered laws of robotics.

Security problems

A single AI agent is relatively easy to control; but a million AI agents acting interactively is considerably harder, even with controller agents. To this must be added the possibility that some agents erroneously reinforce the errors of other agents. In other words, a chain of joint errors may arise from the association. A potential cascade of errors in a highly accelerated process, with no time for pause.

On the other hand, there are the growing security problems. The interconnection of multiple agents creates additional points susceptible to attack. Each agent may have its own malicious agent ready to impersonate it and infiltrate the network of AI agents. I do not know whether the possibility of the "double agent," as in spy networks, applies here. In any case, this makes stronger authentication procedures and continuous oversight critical requirements. Human oversight, presumably. Its promoters do not speak of replacing humans, but of a different collaboration between humans and AI, with creativity, ethical judgments, empathy and the responsibility of oversight resting with the former.

Finally, the socioeconomic consequences that could stem from the society of AI agents. It seems clear that the demand for human labor is lower and of a kind that combines technical specialization with a notable strategic vision. But, above all, it will be lower.

Legal expert specialized in "digital constitutions"

Techno-optimists tend to answer this question by saying that new jobs will be created, such as the designer of AI agent ecosystems, the supervisor or the legal expert specialized in "digital constitutions" who lays the foundations for these AI organizations or who litigates over responsibility for the negative consequences of automatically made decisions. A response about the effects on employment that at the same time sounds understandable - yes, there will be new jobs - and unconvincing, since the jobs created will be far fewer.

To this must be added the great inequality that will arise between societies and large companies with access to enormous and diverse ecosystems of AI agents, with a major leap in productivity, and other societies and smaller companies with less access or access subordinated to such ecosystems. Questions that lead us to the how of the distribution of the returns from the sharp increase in productivity and to the how of the power concentrated over such ecosystems.

In less than ten years, the relationship with a single AI agent will be more obsolete than a modem internet connection. We will interact with ecosystems of agents, in which each agent has accumulated years of specialization. It could be said that, in some way, they will have become professionalized. A society running on an invisible workforce of expert professionals working in coordination. The great complexity will likewise be invisible. Everything will seem easier: an illusion of ease? The definitive incorporation of the techno-solutionism denounced by Mozorov?

The society of AI agents is here. A society that makes decisions in an organized and immediate way. Will the decision of how to organize and associate AI agents, and how to associate with these societies of AI agents, rest with humans?

Javier Callejo

Author

Javier Callejo

Professor of Sociology at the National University of Distance Education (UNED). Degrees in Journalism and Law

The latest
From Huang to Botín, from Zara to Byron: corporate responsibility in global capitalism

Keep reading