AI Agents: What They Are Once You Ignore the Hype
A loop around a model that can take actions. Powerful, and easy to oversell.
"Agent" is the word the industry reached for when a chatbot stopped being enough. It gets thrown around loosely enough now that it's worth pinning down what actually separates an agent from a plain model call — and why the gap between a demo agent and a reliable one is so much wider than it looks.
What actually makes something an agent
A plain LLM call is one shot: you ask, it answers, done. An agent wraps that model in a loop and gives it two things it didn't have — the ability to take actions in the world (call an API, run code, search, write a file) and the ability to decide what to do next based on what happened.
So the loop is: the model looks at the goal and the current state, picks an action, something executes that action, the result comes back, and the model looks again. Repeat until the goal is met or it gives up. That's the whole idea — a model that can act, observe the outcome, and choose the next step. Everything with "agent" in the name is some version of that loop.
Why the loop changes everything
The loop is what makes agents genuinely useful. A one-shot model can tell you how to fix a bug. An agent can read the file, make the change, run the tests, see them fail, and try again. Tasks that need several steps, where later steps depend on what earlier ones discovered, are exactly what a single call can't do and a loop can.
That's the promise, and it's real. It's also where the trouble starts.
Why reliable agents are hard
Here's the part the impressive demos skip. Every time through the loop, the model can be wrong — and now being wrong doesn't just produce a bad sentence, it produces a bad action. It calls the wrong API, deletes the wrong file, misreads a result and confidently builds three more steps on top of the mistake. Errors don't just happen; they compound. A model that's right 95% of the time on a single step is right about 60% of the time across a ten-step chain, and that math is brutal for anything you actually depend on.
So the real engineering of agents isn't the clever prompt that makes the demo work. It's the unglamorous scaffolding around the loop: limiting which actions are allowed, checking results before trusting them, knowing when to stop, and putting a human in front of anything irreversible. The demo is the easy part. The reliability is the work nobody films.
Where they actually fit
Agents earn their keep on tasks that are genuinely multi-step but also verifiable — where you can check whether each action worked. Coding is the clearest example: the tests tell you if the change was good. Research, data gathering, workflow automation over well-defined tools — all reasonable homes. Where they struggle is anything high-stakes and hard to verify, where a compounding error is expensive and you can't cheaply tell it happened until it's too late.
The one-line version I'd give a colleague: an agent is a model in a loop, with hands. That makes it far more capable than a chatbot and far more dangerous — and most of the actual work is making sure the hands can't do too much damage when the model, inevitably, gets a step wrong.
