AI Agents: 2025 Marks a New Era for Artificial Intelligence

An article by Flavien Chervet

02/19/2025

In 2022 the world discovered a marvelous gadget, ChatGPT. But let's not be fooled: the revolution is not there. Chatbots built on generative AI are only the sandbox of a far more transformative phenomenon. In 2025, here we are. Generative AI is reaching maturity and a new era is opening: the assistant becomes autonomous, the tool becomes active, the technology becomes a colleague. In 2025, AI becomes an agent.

From chatbots to agents: a concrete case to understand

The year is 2021. Generative AI has not yet swept across the world. René does not know that his work is about to change shape. René is a consultant in innovation funding. The heart of his job is to talk with entrepreneurs, identify the innovative aspects of their projects, and write up files that explain them to the administration in order to secure funding aid, such as the well-known "research tax credits."

He does everything by hand. The one-on-one meetings, the note-taking, the analysis of the entrepreneur's documentation, the identification of the innovative elements, the writing of the files, the sending of the files. On average, building one file takes him a full 4 days.

And then in 2022, generative AI arrived. René is a geek, so René gave it a try. And René was quite impressed! After taming the beast and learning to speak to it properly by "prompting" the right way, he was able to make his life easier. He still handles the one-on-one meetings. But it's his AI that takes the notes automatically. Then he runs the notes and the documentation through another AI he designed to produce a summary and identify the innovative aspects of the project. Then, when he writes the file, he often asks the AI to draft certain paragraphs, to structure the information, or to improve the style. On average he saves one day on the creation of the file, and what's more, it's of better quality!

And here we are in 2025. René is still a geek. And René has been hearing about "AI agents" everywhere for a few months now. So he wants to give it a try. His goal is simple: fully automate the heart of his job. René wants his agent to talk with the entrepreneur, asking all the relevant questions. Then he wants his agent to retrieve the entrepreneur's documentation by digging through his emails and his folders. Then he wants his agent to analyze the whole thing to identify all the relevant information. Then he wants his agent to write the file for the administration, send it to a second "critic" agent that identifies the weak points, then to a third agent that improves it based on the critic agent's feedback. Then he wants his agent to send it to the administration.

For now, René can't yet put all the building blocks in place efficiently. But his agent is already able to analyze the entire documentation and write a quality file fully on its own. As a result, he manages to handle a file in less than 2 days!

The next building blocks will surely come soon as the technology matures. So René is starting to develop a new strategic consulting offering to help entrepreneurs steer their business model toward more innovation. But he's already thinking about an agent that will do that autonomously!

So, what is an agent?

"Autonomy": that's the key word. Today's AIs are only able to respond to a request and can therefore automate only simple, superficial tasks. They are far more thinking assistants than true colleagues able to create value on their own. But with the advances in the technology, it becomes possible to do better, much better. The concept of "agent" emerged to describe AI systems capable, on one hand, of planning and chaining together many actions, and on the other, of handling "tools" that let them act on the digital or physical world.

Planning and chaining actions

In order to chain several digital actions or thoughts together, you build a "scaffolding" that represents the process you want the agent to carry out, then simply make several successive calls to a generative AI such as GPT, o1, or Mistral. The point is, at each step, to use the AI's previous answers as context for the request of the current step. For example, you can build an article-writing agent with the scaffolding:

Step 1: from the user's topic, propose an original angle

Step 2: from this angle, propose an outline

Step 3: from this outline, write the first part of the article

Step 4: from the outline and the first part, write a coherent second part

Step 5: from the outline and the first 2 parts, write a coherent third part

Step 6: review the whole article, improve the style, and remove redundancies

You can even generalize the approach so the agent can respond to any request, with a scaffolding of the kind:

Step 1: from the user's request, define the tasks to be carried out to obtain a quality result

Step 2: execute the first task

Step 3: based on the result of the first task, check that the task list is still relevant and modify it if it isn't

Step 4: execute the next task

...

Then you keep looping until all the tasks are finished

In the open source code of the agent "Claude-engineer," for instance, you find the instructions (translated into French by the author):

"You are in automatic mode!!!

When you are in automatic mode:

Set yourself clear and achievable goals based on the user's request.

Handle these goals one by one, using the available tools if needed.

..."

The agent thus becomes able to plan and build for itself the scaffolding specific to the current request. The software AutoGPT and BabyAGI are other open source examples of implementations of this kind of architecture. As of today, well-built specialized agents can already be very effective. Generalist agents that plan their own tasks, on the other hand, are not yet robust enough to be used seriously. They sometimes drift gradually away from the initial goal, get lost in unimportant subtasks, or even loop endlessly on a task they aren't capable of completing. These problems come from the limits of the underlying AI models, and improving the latter should make agents more robust in the future. We already see a significant improvement with the latest so-called "reasoning" models such as o1, DeepSeek-R1, or Grok3, which are far more effective at planning tasks.

Tool handling

Just as a human needs a calculator to do arithmetic, a web browser to look for information, or AirBnB to book a place to stay, we can give agents the ability to use external tools to carry out the tasks they will have planned. For example, we can let them run code, access a database or a CRM, browse the Internet, use a specific service like AirBnB (thanks to its API), or use another specialized agent! Thanks to connected objects, we can even give them the ability to act on the physical world.

One of the major areas of innovation right now is "computer use." The idea is to give an agent the ability to understand a computer's graphical interfaces, and to click, scroll, or type text to carry out actions the way a human would. With such capabilities, an AI agent could perform every task a human can perform on a computer, even when they require logging into a portal. The automation potential is immense. The startup Anthropic, which publishes the AI solution Claude, was the first to release a computer use demo in 2024. In early 2025, OpenAI, the company behind ChatGPT, followed suit with its agent Operator, which integrates this feature. However, neither solution is robust enough today to be used in production: the probability that the agent makes mistakes or gets lost along the way is too high. But performance is improving fast! The French startup H Company has, moreover, announced that its agent, Runner H, expected to launch in 2025, reaches human-level robustness in this domain.

These two simple ideas, planning to chain actions together and the use of tools, completely change the game when it comes to the capabilities of AI systems. Imagine the limits a human would run into, even a particularly intelligent one, if all they could do was answer off the cuff, without reflection, planning, or the use of tools. Those are the limits that had been imposed on AI systems like ChatGPT or Claude until now. In 2025, we are entering a new era for AI. By giving them these new capabilities, we unlock a large part of their potential, until now untapped. OpenAI, for example, recently unveiled Deep Research, an agent able, fully on its own, to search for information on the Internet in a highly targeted way, across several dozen source sites, and to write good-quality reports from the information collected. It was tested on the hardest of knowledge evaluations, the "Humanity's Last Exam." Whereas o3-mini, OpenAI's best AI, only reaches a score of 14%, the Deep Research agent reaches 26% (the best score to date)!

Starting in 2025, improving AI systems no longer consists solely of improving the underlying models, but also of creating software architectures in which those models are put to good use effectively. Having a full head is not enough. You need a well-made head!