Suzanne:

Someone asked me a plain question this week: what is an agent? Not agentic AI, not the architecture underneath it, just the word everyone has started using as though its meaning were already settled. It is the right question, and it deserves an answer before all the vocabulary that usually piles up around it. Here is mine.

An agent is an AI you give a goal to, one that then goes off and does the work of reaching it rather than waiting for your next question, so that where a chatbot is something you talk to, an agent is something you hand a job to. As I write this, one is doing exactly that on the other side of my desk. I gave it a time-tracking export from our accounting system, one of those files that is technically data and functionally useless, the kind that costs someone an afternoon just to make sense of, and asked it to turn the thing into a database I could actually analyze. It is pulling the file apart, repairing the formatting, and loading it while I sit here writing about it, and I did none of that step by step; I told it what I wanted and left the room. That is the whole distinction in one small scene, not a chatbot returning an answer but an agent finishing a task.

We wrote a while ago, in “Who Owns the Output?”, about the human end of this arrangement, the fact that responsibility for anything an AI produces stays with the person who signs their name to it, no matter how capable the machine becomes. This is the other end. Before we can talk sensibly about keeping a person in charge, it helps to be clear about what, exactly, is being kept in charge of.

Most people’s picture of AI is still the chatbot, and for good reason: it is what they have actually touched. You type something, it types back. The exchange begins and ends in the box. That back-and-forth is what people mean by generative AI: you prompt it, it generates a response, and that is the whole transaction. It will follow a conversation perfectly well, but it does not know what you are ultimately trying to accomplish, it cannot reach past the window it lives in, and it does nothing at all until you prompt it again.

Within those limits it is genuinely, sometimes startlingly, useful. My favorite recent proof came from my father, who is 88. We took him to France, which meant he missed one of his meetings, and he thought it could have gone better in his absence: the discussion had wandered off the agenda and doubled back on itself for the better part of an hour. So he took the transcript, dropped it into an AI tool, and watched it lay the conversation out in order, separate what had actually been decided from what had merely been said, and hand him a clean list of action items in a matter of seconds. He was delighted, mostly at the time it had saved him, and I was delighted that he had tried it at all. That is a chatbot doing precisely what it is good at, taking one thing in and handing one useful thing back, which is real value and also, just as plainly, the shallow end of the pool.

The agent starts where that leaves off. The easiest way I have found to feel the difference is to picture how you would deal with a contractor. You do not stand over a good contractor dictating each motion. You give them an outcome, renovate this kitchen, along with a budget and a few boundaries, and you let them plan the work, buy the materials, and come back to you for the choices that genuinely need you. You are still in charge; you are simply no longer laying every tile. That move, from answering questions to carrying out goals, is what people mean by agentic AI.

An agent is meant to be handed a job in that spirit, and to manage it, it is given a few things an ordinary chatbot never had: a goal instead of a single instruction, a set of real tools it can operate rather than merely describe, a memory that carries across the whole task so that what it learns at the third step still counts at the twelfth, and permission to act within limits you draw in advance. What makes it an agent, though, is less any one of those ingredients than the way it uses them, which is in a loop rather than a single reply. It takes stock of the situation, decides on a step, takes it, looks at what happened, and goes around again, correcting as it learns, until either the job is done or it hits something it should not settle on its own. Where a chatbot answers once and then waits, an agent keeps working, the way a competent person would, which is why it can carry a task through a dozen moves instead of one.

If I had to name the single capability that matters most, it would be the tools. A chatbot can write you a flawless set of instructions for reconciling a stack of contracts; an agent opens the systems and reconciles them. The distance between describing the work and doing it is most of what separates this era of AI from the last.

There is a Seinfeld episode that shows the frustrations at opposite ends of that balance, and it is worth ninety seconds of your day for the laugh alone. Jerry hires a contractor named Conrad to redo his kitchen, and Conrad turns out to be maddening in his deference, checking on every last detail until Jerry, worn down, tells him to just decide things himself. Freed of any guidance at all, Conrad builds an enormous, preposterous addition that nobody asked for. The joke doubles as a design principle. A contractor who needs you for every decision is useless, and one with no boundaries whatsoever builds you something you never wanted. The work lies in setting the guardrails, and then trusting what is inside them to make the small calls.

Adeline:

The Conrad joke lands a little differently from where I sit. I have only been working in this field for a few months, but it has been long enough to see that same instinct already written into law before I arrived. The whole architecture of the EU’s AI Act starts from the assumption that you set the boundary before you switch the system on, not after it has already built you the addition nobody asked for. It sorts uses by how much room they are given to decide things on their own, and it asks for the tightest limits exactly where the stakes are highest, not because agents are dangerous by nature but because a boundary discovered after the fact is really no boundary at all, just an expensive lesson.

That is the piece I would add to the contractor picture. The budget and the boundaries are not the paperwork you get through before the interesting part starts; they are the interesting part. Decide them well and the agent’s freedom inside them is what makes it worth having in the first place. Decide them badly, or not at all, and you end up relitigating after the fact what should have been settled in advance, which is a much slower and more expensive way to arrive at the same boundary.

Suzanne:

Adeline has it exactly, and it is the whole point: determine the guardrails before the work begins, and everything inside them stops being a leap of faith and starts being useful. It also changes what you find yourself handing over. What I actually reach for day to day is rarely a question and an answer at all. I built and deployed a small iPhone app for a trip to France, about forty-five minutes of work, so we could split expenses the particular way I wanted instead of paying for the usual subscription. I have an agent working through my team’s sprint history, open roles, and time records to tell me where we are genuinely short-staffed. And for this blog, one agent pulls and sorts the news I draw on while another lives in the shared folder my daughter and I write in, reading and editing the files directly.

Plenty of the rest is closer to a very fast assistant than a true agent, decks and proposals and plans drafted in minutes and then reworked by me, and, honestly, almost everything else besides. I do not really work without it anymore, which is exactly why it is worth keeping straight in your head which of these you are using at any moment: the thing you talk to, or the thing you hand a job to.

Adeline:

On my side of the Atlantic, the day-to-day picture looks less like a tidy list of named agents and more like a spread, because everyone around me is standing somewhere different on it. Some colleagues still treat AI as a chatbot in exactly the sense Suzanne describes, a faster way to get a first draft or a second opinion, and are perfectly right to use it that way for now. Others have already handed over the kind of multi-step, tool-using work she is describing here, and would tell you that going back to doing those steps by hand feels almost willfully inefficient.

I do not think that gap says much about the technology itself. It says more about how much of any given job can be broken into steps clear enough to hand off, and how comfortable each person is deciding, in advance, which of those steps still needs them. That, more than which model or which interface, is the variable that actually predicts who is using a chatbot and who has quietly started using an agent.

Suzanne:

Adeline’s point is the one I would underline. Whether you are using a chatbot or an agent turns out to say less about the software than about how well you understand your own work: which steps can be handed off, and which ones still need you. That, at least, is the what. The harder questions I will save for the second part of this: how these systems actually work underneath, why that becomes so important the moment your work has to be repeated or audited, and what happens when your agent starts dealing directly with someone else’s. That is where the genuinely interesting part begins.

Suzanne writes from the US; Adeline writes from the EU. Neither of us is handing the pen to the machine.

Sources & further reading

Earlier in this series: “Who Owns the Output?” (Transatlantic Intelligence)

And because it’s worth it: Seinfeld, “The Nap” — Jerry’s new kitchen cabinets (clip)