Beyond Chatbots: Let’s Talk About Agentic AI

Until recently, we used Generative AI in an essentially passive way: we sent a command (a prompt), read the answer and, if needed, asked for another one. That was it.

Today, the real technological revolution is taking a very different direction: Agentic AI. Unlike a traditional model, an AI agent does more than respond. It takes on a goal, identifies and plans the steps needed to achieve it, uses external tools on its own and can potentially correct its mistakes until the task is complete.

In short, artificial intelligence is moving from being a simple production "assistant" to becoming a genuinely autonomous collaborator, capable of completing complex workflows with only minimal supervision.

The table below compares the two approaches.

When looking at its origins, it is important to distinguish this technology's conceptual roots, which go back decades, from its much more recent emergence and commercial definition.

🏛️ The Conceptual Origins (1950–2001)

The idea of a machine capable of acting as an "autonomous agent" goes back to the founding figures of computer science:

• The visionary idea (1950–1959): As early as 1950, Alan Turing imagined machines that could not only calculate, but also make decisions. In 1959, scientist Oliver Selfridge presented his "Pandemonium" model, describing intelligence as the combined work of little "demons" (specialised software, not the kind we meet at night :D). Working together, they anticipated today's multi-agent systems by sixty years.
• The prophecy (2001): Tim Berners-Lee, the inventor of the World Wide Web, theorised in his essay on the Semantic Web that "intelligent software agents" could browse the web, book flights or manage calendars on users' behalf without human supervision. Of course, the technology to make that happen was missing at the time: language models (LLMs), for a start.

🚀 The ReAct method: reason and act (2022)

In October 2022, a group of researchers from Princeton and Google Research presented ReAct, one of the most important methods to contribute to the development of agents based on language models. The model alternates reasoning with actions in an external environment, using the results to decide what to do next.

• What happened: The researchers published the landmark paper "ReAct: Synergizing Reasoning and Acting in Language Models".
• The breakthrough: ReAct showed that a language model does not have to stop at producing a single answer. It can alternate reasoning — Reasoning, as its friends call it — with an external action — Acting, to its friends too 😁. That action might be an API call, a Google search or reading an email. After each action, the model observes the result and decides how to proceed. This cycle lets it tackle a task step by step and reconsider its approach when necessary.

🏷️ The Birth of the Term "Agentic" (2024)

Although frameworks such as LangChain and AutoGen were already emerging in 2023, the scientific community was still arguing over what could truly be called an "agent" and what could not.

• Who coined it: In March 2024, renowned AI scientist and pioneer Andrew Ng (co-founder of Coursera and Google Brain, in case you're wondering) introduced and popularised the term Agentic AI.
• Why: Andrew Ng chose the adjective agentic (borrowed from psychology, where it refers to a person's ability to exercise control over their own life) to move beyond the binary debate. Instead of arguing about whether a piece of software is "a 100% pure agent", he proposed looking at how agentic an AI's behaviour is: its ability to plan, reflect and use tools.

Here's a handy little diagram summarising the main milestones:

Enlarge diagram

As you've probably gathered by now, Agentic AI emerged to address a frustrating and rather tedious structural limitation of traditional Generative AI:

• The need for constant prompting: With classic chatbots, such as the first version of ChatGPT, the user has to play "director" and enter one prompt after another: "Search for this, now summarise it, now draft an email, now send it." Agentic AI aims to let users state the final goal — "Do this research and send the results to xyz" — while the AI handles the tedious intermediate steps on its own.
• Overcoming hallucinations: Systems based on agentic cycles, such as reflection or the use of sandboxes, can drastically reduce model errors because the AI can check facts before responding, test code or query databases to validate its answers.

As we have seen, Agentic AI traces its primordial origins back to the 1950s, only to effectively take off from 2022 to the present day.
It takes charge of a goal, plans the steps, uses tools, verifies the results, and re-evaluates its own actions.
It requires minimal supervision and drastically reduces the errors (hallucinations) of its predecessors.
The future of Agentic AI promises to be extremely dynamic and paves the way for previously unimaginable applications.
It could certainly be worth getting to know it better and starting to leverage it to develop our ideas.
But an important question is: when is it truly worth using?
In the next installments of this Series, we will understand when to choose it, when to do without it, and what the main patterns are behind this fantastic technology.
Until next time and... Stay tuned!!