San Francisco – May 20, 2026

By Emilia Pasquier, CEO of Swissnex in San Francisco

Part of my work is attending conferences on new trends and technologies — which means I often hear things before I fully understand them. I vividly remember the first time I heard someone talk about AI agents. It was a pitch night in March 2023, during what would be remembered as the rainiest winter California had seen in decades. I had just moved from Switzerland to the Bay Area, had been promised endless sunshine, and ended up wearing plastic boots and a rain poncho for weeks.

One evening, I fought the rain to attend an event in San Francisco — a pitch night featuring former Big Tech workers turned founders. You might remember that early 2023 came with a significant wave of tech layoffs. The Bay Area was wallowing in its latest doom era. While the rain was falling, major tech companies were cutting jobs. The hypotheses ran the gamut of possibilities: Was this the end of Silicon Valley? A correction after pandemic-era over-hiring, when every company had gorged on talent to serve a world that had moved entirely online? Or something else entirely?

When an industry cuts jobs, there’s usually a lot of negative spin. But the Bay Area doesn’t follow that script. When there are layoffs, investors see a talent pool ripe to become founders. And that’s exactly what was happening that March night three years ago. ChatGPT was released to the public just a few months earlier, and at the same time, there was a wave of brilliant, unemployed people ready to build something using that new tech.

That evening, for the first time, I saw someone demonstrate that a model could go beyond answering questions almost like a human — that it could actually run tasks autonomously. I was blown away. I spent the whole walk home, soaking wet, imagining the applications: a personal assistant booking my flights, an autonomous system ordering my groceries every week, something that would just handle things. In retrospect, I wasn’t thinking grand enough.

And what I definitely didn’t anticipate was that before an agent can work for you, you have to raise it.

So How Do You Raise an Agent?

Fast forward to 2026. We all hear that agents will revolutionize work, that the next generation of workers will manage agents instead of people. And yet, when you attend a tech conference today, AI engineers sound less like technologists and more like kindergarten teachers.

Here is a paraphrased version of something I have heard a couple of times at recent conferences:

“You can’t just let your agent run from day one. You have to give it autonomy step by step — like a child. Start by putting them in the sandbox. At first, you watch constantly: don’t eat the shovel, don’t put sand in your mouth. Once you trust them, you step back a little. Then you walk them to the end of the block to make sure they can navigate the street. Eventually you let them go to the neighbor’s house — but they have to be back by 5pm. Then 10pm. And finally: the sleepover.”

It is striking to hear AI engineers use this kind of language. I want to dig into this analogy — because I think it reveals that what is at stake with agents goes far beyond building an intelligent system. It is about creating an autonomous one, and this forces us to ask harder questions: how do we define autonomy? And are we ready to delegate it to machines?

The Drinking Age Problem

Raising a child is, at its core, a project to teach autonomy. The word itself comes from the Greek autonomos — auto (self) + nomos (to govern). And autonomy, as a concept, carries enormous philosophical weight. Across traditions — Kantian ethics, developmental psychology, political philosophy — becoming autonomous implies for instance the capacity for decision-making, moral reasoning, and self-reflection.

As a society, we translate these ideas, among other things, into law. We decide at what age a person is mature enough to make certain decisions and bear the consequences of their actions. And we don’t agree: Switzerland lets you buy beer at 16; in the U.S. you have to wait until 21. Switzerland won’t let you drive until 18; the U.S. allows it at 16. The lines are drawn differently, but the underlying question is the same: when is someone ready?

Sitting in that AI conference, listening to engineers talk about agentic autonomous systems while explicitly invoking the language of child-rearing, I found myself wondering: what is the drinking age for an AI agent? At what point can we let them drive?

The child analogy, I think, is telling us something important: we are moving toward a world where models will take on a real share of human decisions and actions. The question of when a model is sufficiently autonomous to be trusted with that responsibility is one we have barely begun to answer.

From Faster to Wiser to More Autonomous

The recent history of artificial intelligence has moved through distinct phases: building large language models, making them faster, more coherent, and less energy-intensive. Each phase had a clear metric for success.

But the child analogy suggests we have entered a different phase entirely. We are no longer just building tools that assist decision-making. We are creating models that can take on parts of our decision-making, act on our behalf, and operate with a degree of independence in the world. We are delegating a slice of our autonomy to them.

Which means the next critical question isn’t how powerful is this model? It’s how do we decide when it’s ready? Who draws the line? In other words: who sets the drinking age for agents, and what does their driving test look like?

I don’t think we’ve figured that out yet. And I don’t think these questions should be answered by engineers alone, however thoughtful they may be. They need to be sorted out across disciplines — psychology, philosophy, political science, sociology — by people who have spent their career looking into what it means to act responsibly in the world.

To raise a good human takes more than technical skill. We gave that responsibility to parents, to communities, to legal systems built over centuries.

Building, or raising, good AI agents demands the same breadth of thinking. The goal isn’t just performant agents — it’s agents that can leave the sandbox with a solid ethical framework before they start ordering drinks on our behalf.

What is your experience with training AI agents and what do you think is critical to making sure the next step of their “childhood” is on track? Please share your thoughts on the Linkedin post.

This is part of a newsletter published on LinkedIn. Subscribe here to get alerted about new articles.