What is an agent?
Everyone is talking about agents. Very few people can tell you what one is. Here's the short version, and a playground where you build one yourself.
Let me start with the question I get asked the most: what is an agent?
Not an agent from The Matrix. I know, that would make for a way cooler blog post. The real thing is simpler, and once it clicks you'll see it everywhere: an agent is a language model using tools in a loop.
Honestly, you already know what an agent is, because we've had them forever. Think of a travel agent, back in the day. You'd walk in and say "I want a beach week in August, keep it cheap." That's your intent, fuzzy and human. They'd ask a couple of questions (how far? window or aisle? kids or no kids?), then work their tools: the booking system, the phone, the thick binder of packages. After each call they'd check what came back against what you asked for, adjust, and keep going until you had an itinerary in your hand.
The travel agent at work: intent in, questions, tools in sequence, itinerary out.
That's the whole shape of it. Swap the human for a language model and you have an AI agent. Three parts:
The three parts
A model.
The brain. What made agents take off is that models got good at deciding when and how to use tools, not just at writing text.
A harness.
The software around the model. It shows the model which tools exist, runs whichever one the model picks, and hands back the result. The model never touches your files or the internet directly. It asks, the harness does.
The loop.
After every result the model checks its own progress. Did that work? What's next? It keeps going until the job is done or it needs to ask you something. A chatbot answers and stops. A workflow follows steps a developer wrote in advance. An agent decides its own next step, every step. That loop is also why an agent can chase a goal for hours or even days, but that's a post of its own.
One thing about trust: you decide which tools an agent gets, the same way you'd onboard a new hire with the right permissions instead of the master key. An agent can't take a destructive action it doesn't have access to. Guardrails are a feature.
Enough reading. Build one.
The playground needs JavaScript. The essay doesn't.
The shelf
Drag a tool onto the agent, or just tap it.
Guardrails are a feature.
The agent. Tools dock around the harness ring.
Harness transcript
Every step, logged. The log is the whole trick.
What about the physical world? Think of a robot vacuum. It has tools (wheels, brushes, sensors), it has context (a map of your living room), and it pursues a goal. What it's missing is the flexible brain that can handle "actually, skip the rug, the cat is asleep on it." Pair a body like that with an agent brain and the line between computer work and real-world work starts to blur. The robotics labs are already on it.
And one last thing, the part I actually want you to take away. The obvious move is to hand agents your existing workflows. The better move is to ask why a workflow exists at all. Hand the agent the problem, not the process. What's upstream? What root cause, if fixed, means nobody, human or agent, ever has to do this again?
Optimizing a workflow gives you the same work, faster. Deleting a workflow gives you the time back entirely.
Those are not the same prize.
So that's my honest answer. An agent is a model, a harness, tools, and a loop. But what it really is, is the thing that takes problems you'd filed under "someday, if I ever get the time" and moves them to "this week."
Sources: Building effective agents · "tools in a loop" · Writing tools for agents · Long-running harnesses · ReAct · Toolformer · Agent survey · CoALA · Tool use docs · Claude for Chrome · Gemini Robotics 1.5