Booking a trip by comparing hundreds of offers, writing a report from multiple documents, analyzing medical data, or automatically correcting a computer program: these tasks require thought, method, and a variety of skills. “Agent AI” now promises to accomplish them autonomously, orchestrating the necessary operations, using tools, and correcting its own errors. However, current agent AI does not yet encompass the full richness of the “autonomous agent” concept as it was developed in previous decades.

Gartner has declared 2026 the year of “AI agents .” These systems go beyond simply improving chatbots. OpenClaw ‘s AI agents are already capable of communicating with each other and performing complex tasks with limited human oversight. For businesses in industry , government , and healthcare , the promise is more flexible automation than traditional software, capable of adapting to diverse situations rather than applying predefined rules.

However, behind the apparent novelty lies a longer history. Agentic AI is part of decades of research on autonomous agents and multi-agent systems . What is changing today are the tools, particularly the large language models and their ability to interact more naturally with humans.

From text generation to action

Conversational models, such as ChatGPT, Gemini, or Claude, impress with their ability to summarize or write complex texts. However, taken individually, they remain essentially reactive: they produce a response based on a query. An autonomous agent goes further. It can analyze a request, plan a sequence of operations, use external tools (search engines, databases, software), evaluate the result, and adjust its strategy if necessary.

Whereas a language model is limited to writing a computer program, an agent can execute it in a secure environment, observe any errors, correct the code, and then test it again. In short, AI agents don’t just talk, they act.

The shift from text generation to action transforms the very nature of software. While a program follows precisely defined instructions, an autonomous agent can dynamically adapt its decisions based on context, results obtained, and objectives. It doesn’t necessarily replace humans, but it does alter the distribution of tasks between supervision and execution.

Promises and risks

This development opens up considerable possibilities. Within organizations, agents can automate laborious business processes . In industry, they can coordinate complex software systems . In the medical field, they can analyze records, search for relevant publications, and provide summaries to assist physicians . But these promises come with risks.

Indeed, current models can produce inaccurate information, the infamous hallucinations, and are likely to reproduce biases present in their training data. If agents are confined to an assistant role, these limitations are already problematic; they become critical when they concern systems capable of acting on technical infrastructures, notably by executing system commands, manipulating files, or sending network requests.

The issue of AI agents is therefore not only technical: it is also legal, economic and societal. It touches on the transformation of skilled work and the governance of IT systems .

A historical lineage

The idea of ​​an autonomous agent did not originate with language models. It dates back to the very origins of artificial intelligence. In 1956, at the founding conference in Dartmouth (in the northeastern United States), one of its organizers, Marvin Minsky , already defined AI as the design of programs capable of performing tasks that mobilize so-called intelligent capacities such as understanding, learning, reasoning, or deciding.

From the 1980s onward, the concept of the ”  intelligent agent  ” became central. An ”  agent  ” was then defined as a program capable of perceiving its environment, making decisions, and acting to achieve objectives. Early on, researchers developed the field of multi-agent systems: organized sets of autonomous programs that interact within the same digital environment. The goal was to understand how these entities could coordinate, cooperate, or compete to solve complex problems.

Several landmark projects concretely illustrate this approach. The HEARSAY-II system is based on a “blackboard” model. Several specialized modules for language recognition, analysis, and interpretation contribute to speech comprehension by sharing their hypotheses in a structured common space. The Contract Net Protocol offers a mechanism inspired by calls for tenders: to perform a task, an agent issues a request for proposals, other agents offer their services, and the most competent are awarded the contract. In other words, coordination between agents has been at the heart of AI for several decades.

A still largely untapped reservoir of ideas

While the concept of an agent is not new, “agentic AI” is gaining traction among non-specialists today due to the central role played by large language models. Although lacking causal understanding and comprehension of the physical world, these models provide agents with linguistic capabilities and a form of statistical “common sense” that facilitates interaction with humans and the interpretation of complex instructions in natural language.

However, current agentic AI does not yet encompass the full richness of the autonomous agent concept as it was developed in previous decades. In practice, it still most often relies on a sequence of actions where each step is planned and ordered in advance. The work carried out since the 1990s on multi-agent systems, focusing on cooperation , negotiation , task allocation , and collective adaptation, offers a reservoir of ideas that remain largely untapped.

Integrating these mechanisms with the capabilities of large models opens up new perspectives for tomorrow  : agents capable not only of executing a plan but of organizing themselves collectively, specializing and adapting to complex environments.

Agentic AI thus represents a new stage rather than a radical break. It combines the theoretical legacy of multi-agent systems with the recent power of generative models. Understanding this historical lineage allows us to move beyond the hype. Agentic AI represents an attempt to transform predictive models into systems capable of acting, planning, and perhaps, in the future, organizing themselves collectively in complex environments.

Author Bio: Maxime Morge is Professor of Computer Science at Claude Bernard University Lyon 1