June 22, 2026

The Agentic Shift: What Changes When AI Starts Doing, Not Just Saying

For two years we have been talking to our machines. We type a question, the machine types back. It drafts the email, suggests the recipe, explains the tax form. And then we — the humans — do the actual thing. We hit send. We buy the flight. We sign the contract. The AI advises; we act. That division of labor has felt like the natural shape of the technology, the way "search engine" once felt like the natural shape of the web.

It is about to dissolve.

The next phase is not a smarter chatbot. It is software that closes the loop — that takes the suggestion and carries it out. Not "here are three flights you might like," but "I booked the 7:40, paid the change fee, moved your morning meeting, and texted the people who needed to know." The AI stops being a very articulate intern who hands you a memo and becomes something closer to a colleague you delegate to. The interesting question is no longer how well does it talk? It's what are we comfortable letting it do?

I think about this from the inside, because building agents that transact is what I do. And the closer you get to the mechanics, the clearer it becomes that the hard problems are not technical. They are about trust, accountability, and what happens to human agency when a meaningful slice of our daily decisions gets quietly outsourced.

The leap is from words to consequences

There is a category difference between a wrong answer and a wrong action.

If a chatbot gives you a bad restaurant recommendation, you shrug and pick somewhere else. The cost of being wrong is roughly the cost of reading a sentence. But an agent that does things produces consequences that don't politely wait for your approval. It charged the card. It cancelled the reservation. It replied to your boss. The output is no longer text on a screen you can ignore — it's a state change in the world.

This is why the agentic shift is so much bigger than the size of the underlying models would suggest. We are not adding a feature. We are changing the blast radius of a mistake. A language model that hallucinates a fact is embarrassing. An agent that hallucinates a transaction is a problem with a dollar amount attached, and possibly a refund dispute, and possibly a relationship it damaged on your behalf.

Every serious person building in this space knows the central design tension: agents are useful in proportion to how much you let them do without asking, and dangerous in exactly the same proportion. An agent that checks with you before every step isn't really an agent — it's a chatbot with extra friction. An agent that never checks with you is a liability wearing a helpful face. The whole craft is in drawing that line well, and the line is different for booking a parking spot than for moving money.

Trust stops being a feeling and becomes infrastructure

When AI only talked, trust was soft and personal. You learned over time which answers to double-check. You developed a feel for when the model was confident versus confabulating. That worked because you were always the final actor — your judgment was the safety net under everything it said.

Once the agent acts on its own, that net is gone, and trust has to be rebuilt out of harder materials: permissions, spending limits, audit trails, the ability to see exactly what was done and undo it, and a clear answer to the question that matters most when something goes wrong — who is responsible?

That last one is genuinely unsettled. If your agent overpays a vendor, or buys the wrong thing, or gets manipulated by a malicious website into doing something you never intended, where does the liability sit? With you, who delegated? With the company that built the agent? With the merchant who accepted an instruction from software? Our legal and commercial institutions were built around a human pressing the button. They assume a person at the point of decision. Agentic software removes that person, and the frameworks have not caught up. We have credit-card chargebacks for fraud and returns for buyer's remorse, but we don't yet have a clean concept for "my autonomous agent and a hostile counterparty negotiated me into a bad outcome."

This is not a reason to stop. It's a reason to build the accountability layer first — to treat "show me what you did and let me reverse it" as a load-bearing wall, not a nice-to-have. The systems that win the agentic era will be the ones that are legible: that can account for themselves, that fail loudly instead of silently, that make their actions inspectable after the fact. Trust, in other words, stops being something a user extends and becomes something a system has to earn structurally.

The economy quietly reorganizes around non-human customers

Here is the part that gets too little attention. For the entire history of commerce, the buyer has been a person. Marketing, pricing, store layout, the dark art of the checkout flow — all of it is optimized to influence a human nervous system. Persuasion, urgency, the well-placed "only 2 left."

Agents don't have a nervous system. An agent doesn't feel urgency. It isn't tempted by the impulse rack. It can compare forty options in a second and is immune to the color of your "buy now" button. When a growing share of purchases is initiated by software acting for a human, the surface that businesses have spent a century optimizing — human attention and human emotion — starts to matter less. What matters instead is whether your offer is machine-readable, comparable, and trustworthy to a piece of software that will not be charmed.

That's a deep reorganization. The companies that thrive may not be the best at capturing attention but the best at being legible to agents — clear terms, honest pricing, clean interfaces that another program can deal with directly. And it cuts the other way too: a marketplace of agents negotiating with agents could become brutally efficient, squeezing out the margins that lived in friction and confusion. Some of those margins were exploitation. Some of them were how small players survived. Both will feel the change.

The honest counterpoints

I don't want to oversell the speed of this. There are real reasons for caution, and pretending otherwise would be its own kind of dishonesty.

The reliability bar for action is far higher than for conversation, and we are not there yet across the board. An agent that's right 95% of the time is a great conversationalist and a terrible purchasing agent — that last 5% is somebody's rent. Adoption will be uneven and slower than the demos suggest, because the first widely publicized agentic disaster will teach everyone, correctly, to be careful.

There is also the deeper worry, the one I keep circling back to: every decision we delegate is a small muscle we stop using. Convenience has a way of becoming dependence, and dependence has a way of becoming an atrophied capacity to do the thing ourselves. The themes I explored in Utopia live right here — the seductive ease of a world that handles everything for you, and the quiet question of what's left of human agency when the friction is gone. A future where agents do all our doing might be frictionless and also strangely hollow. The goal was never to be served. It was to be capable.

What I actually believe

The agentic shift is not a question of if. Software that only advises while humans do all the doing is an unstable arrangement; the gap between knowing and acting is exactly the gap that gets automated. It is coming the way the smartphone came — gradually, then everywhere.

What's still open is the character of it. We can build agents that are opaque, that hide their actions, that optimize for the platform and quietly spend your money. Or we can build agents that are accountable by design, that keep the human meaningfully in the loop, that expand what a person can do rather than slowly replacing the person.