AI Agents: a new stage of automation.

A new grail is driving the ambitions of tech giants: the creation of a new type of software known as “AI agents,” capable of carrying out actions in order to achieve defined objectives. They are distinguished by their ability to reason, learn, adapt, decide, and plan actions autonomously. They are also able to collaborate with one another to coordinate and execute workflows, that is, organized sequences of tasks or actions of varying complexity. They can be deployed across many sectors to automate, streamline, and accelerate business processes.

Within a collaborative workflow in a customer service department, for example, one agent may be programmed to analyze a request. Once the request has been interpreted, it is passed on to another agent responsible for searching for the answer in a knowledge base, while a third agent is tasked with delivering the response to the customer — all without any human intervention.

Use-case scenarios are virtually endless and may concern every level of an organization, whether support, sales, logistics, finance, inventory management, information retrieval, data analysis, strategic analysis, legal or economic monitoring, and so on. In other words, AI agents represent a major evolution in information processing and process automation, with an extremely broad range of applications regardless of industry or function. They can operate wherever repetitive or structured tasks can be automated and can contribute to significant improvements in labor productivity. Hence the enthusiasm they currently generate, in the wake of the ChatGPT wave, both among software publishers and among corporate users.

The capabilities of AI agents are made possible by foundation models and by the multimodal capacities of generative AI, that is, the joint processing of different types of information (text, voice, sound, code, images, video, etc.). The Google Cloud website identifies five essential characteristics specific to AI agents:

• Reasoning: A fundamental cognitive process, reasoning consists in mobilizing logic and available information in order to draw conclusions, produce what logicians call inferences, and solve problems.

• Execution: Execution refers to the ability to act and perform tasks based on decisions, plans, or external information, in order to achieve precise objectives. This presupposes that AI agents can interact with their environment. The action or task may be physical, in the case of embodied AI, or digital — for example sending a message, triggering a process, or modifying a system.

• Observation: An agent always operates within a given context. It must therefore be able to collect relevant information about its immediate environment in order to understand the situation it is in and make the most appropriate decisions possible.

• Planning: Once an objective has been defined, the AI agent must be able to develop a strategic plan to achieve it, select appropriate means, and organize the steps necessary for its realization.

• Collaboration: In increasingly complex and dynamic environments, AI agents must be able to work effectively with humans or with other AI agents, in order to coordinate sequences of actions and achieve a common goal. This collaboration implies the ability to interact, communicate, understand intentions, and integrate the perspectives of other actors, whether human or software-based.

• Self-improvement: Finally, the capacities for self-improvement and adaptation constitute central properties of advanced AI systems based on machine learning. They refer to the ability to learn from experience in order to adjust behavior and improve performance over time.

It is therefore important not to confuse AI agents with AI assistants or chatbots. Whereas AI agents are capable of proactively and autonomously executing complex multi-step tasks and making decisions on that basis, AI assistants are limited to helping users carry out tasks by responding to their requests, while chatbots mainly automate simple conversations based on predefined rules. It is thus the proactive and autonomous character, the capacity to execute complex task sequences, and the ability to learn from experience in order to improve performance that fundamentally distinguish AI agents from other categories of conversational tools.

According to Sam Altman, founder of OpenAI, AI agents would correspond to “level 3” on the evaluation scale of AI systems, following systems capable of reasoning (level 1) and conversational robots such as ChatGPT (level 2). “Level 4,” presented as the next step in the quest for superintelligence and as an almost inevitable horizon of AI development, would consist in designing agents more intelligent than human experts in their domain, capable not only of executing complex tasks but also of innovating or managing the work of an entire organization. In this perspective, current AI agents appear as enhanced versions of existing conversational assistants such as Copilot (Microsoft), Gemini (Google), ChatGPT (OpenAI), or Claude (Anthropic).

Beyond the promises put forward by AI system designers, who announce the imminent ubiquity of agents capable of autonomously carrying out complex tasks like highly skilled employees, or acting as omniscient assistants and advisors made available to all, AI agents nonetheless raise major questions. These concern in particular the security of information systems, the control of how they will be used — whether for beneficial or malicious purposes — as well as the potentially rapid and difficult-to-anticipate effects on employment, insofar as it may become possible to automate at scale, and at lower cost, complex tasks that were previously performed by humans.