If You Don't Try AI Agents You're Incompetent

First generation: the chatbot in a dark room

The first generation of AI, which arrived as a product around 2022 with ChatGPT, is the one made of models experienced by people as chatbots. You go to the site, you talk to this thing, this neural network, and it answers you.

Now, what is it about these models that you have to understand? A model is a text completer. It’s a statistical compression of all the text on the internet, with some adjustments, but it essentially expresses itself in terms of language: text in, text out. Then there are the multimodal ones that also deal with images, but these models are fundamentally neural networks, a bit inside a black box, closed off from the world.

People had a lot of fun, they started playing with them, they saw that they translate, that they write marketing copy, and in particular schoolkids realized they do their homework for them. But they’re chatbots: they take text as input and essentially spit text back out.

Second generation: agents that act

The second generation of AI, the one made of agents, is profoundly different, and it’s built around, or as programmers like to say, on top of, the first generation.

What are these agents? If you have the model and nothing else and you tell it “open a support ticket for me,” or “put eggs on my shopping list, which happens to live in another app,” the model on its own can’t do those things, because it has no access to external resources and services. An agent lets the model, which acts as a language engine for this new second-generation architecture, make decisions based on what you ask it to do. Unlike the first generation, which was model-and-nothing-else, basically shut inside a dark room, this model can now say “let’s do this, let’s do that,” and this and that are services that actually have an effect on the world. Create an event on the calendar, send a bank transfer to someone. This is software built around the model, and it’s an extremely sharp jump, because you go from the chatbot, which is useful and fun and everything you want, to entities that are operational, and profoundly so.

An agent is a bit like an entity that can automate anything that can be digitized. At that point all it’s missing is a body, and at some point we could move to a third stage of AI, the embodied one, where actual robotics comes in, and you realize you finally have something doing things in your place because it does them with a body. But for now we’re in an in-between phase.

Why programmers are in turmoil first

These agents, folks, have exploded in the world of programmers, but they’re coming for everyone. Let me put it very concretely. If you wanted to write code, say to build a site or an app of any kind, and you wanted help from first-generation AI, the models, you had to copy and paste the code into the ChatGPT site or whatever, the model, which is linguistic and has seen a lot of code in the text it was trained on, would try to help you with that code, you’d look at it, go “ah okay, we like this,” and recopy and paste it back where the code actually lives, in the server, in the development environment, whatever it is.

With agents, in particular from Claude Code onward, which is an Anthropic product, and now there’s also the open counterpart called Codex, and open versions too, one is called Open Code, you open a terminal or your development environment and the famous chatbot is actually right in there. And it’s not a simple chatbot that’s purely textual, that just converses. When you tell it to do things, it can read all the files inside the project on its own initiative, without you necessarily explaining where things are. It searches for them, modifies them, asks you, depending on how you configure it, whether that thing can be changed or not, and it writes code directly into the project. This is agentic: it acts, it does things. It’s not a chatbot. A chatbot is nothing next to a system like this.

So for anyone still convinced, “yeah but chatbot here, chatbot there,” no, folks, if you don’t hurry up you’re going to take a lot of slaps. I’m saying this above all to a whole category of characters, business owners, veteran developers who put up resistance. If you don’t hurry up, the twenty-year-old kids will mow you out of the economy, and at that point you’ll deserve it.

It comes for the doctor, the lawyer, the secretary too

These agents do things, and developers are living through it first because they’re adjacent to the technology, the products reach them immediately. But the concept of an agent can be applied to the secretary, the lawyer, the doctor. So does that mean the doctor is no longer needed, the lawyer is no longer needed? No. It means that when the doctor needs a piece of information, this agent is already embedded in patient databases, in pharmaceutical databases, in databases about chemistry, biology, and a thousand other sources, and the patient history. When the doctor asks the agent something, it’s the doctor who knows what to ask and how to evaluate what comes back, but the agent on its own can do anything, including preparing the report as a draft, so the doctor then reviews it. Do I even need to give the lawyer example too? It’s not full automation, it’s heavy, partial, but very heavy automation.

Many in the developer world are having big trouble making this jump, because having worked for twenty or thirty years in a certain way, they now find themselves reinventing and working not so much on the craft of writing code itself, but on managing the delegation made to these machines. That’s the job. You become the bottleneck, as a human being, because you have to control where the risk is highest, and so the craft changes completely. That’s the big difference. Programmers are living it first, but it’s coming for everyone, and it’s monstrous.

Stop talking about chatbots

In a period like this, with a huge new war, tremendous economic uncertainty, social fractures on every side, anything can happen. Go look at this stuff and start experimenting. Stop talking about chatbots, otherwise it’ll pick you off one by one and knock you senseless, because you have no idea what’s happening.

And about the comments that will probably show up, “no Piero, you’re still saying these things? I’ve got the rough loop, I’ve got MCP.” I’m happy for you, well done. But I assure you that, going around and running courses in companies, even multinationals, it’s not true in the slightest that people are aware of what’s happening. On the contrary, you have to help me spread this. So share this video with a friend who’s an AI denier, especially if they’re a manager and a certain age, because those are the hardest to convince.