What Is an AI Agent?
How it differs from a chatbot, and why it's suddenly everywhere
① The line between a chatbot and an agent isn't capability, it's loop structure. A chatbot answers one question and stops; an agent takes a goal and works through multiple steps on its own, with no one in the loop.
② What makes that loop possible is a single capability: function calling (tool use) — the model executes an external program and feeds the result back into its own input to decide the next move.
③ What agents scale up isn't intelligence, it's the total volume of inference requests. One question becomes dozens of API calls, which meaningfully changes the shape of AI infrastructure demand.
① Where the line actually sits between chatbot and agent
The word "agent" has been everywhere since 2025, but ask people to define it and the answers scatter. Some say "a chatbot that understands you better." Others say "an AI employee that works alone." Neither is wrong exactly, but both miss the point.
The line isn't about capability, it's about structure. As covered in how LLMs work, all a model ever does is take input and predict the next token. A chatbot runs that process once and hands the result to a person. An agent wraps that same process in a loop — the model's output stops being "an answer" and becomes "the next action to take," and the result of that action feeds straight back into the model's input.
| Chatbot | Agent | |
|---|---|---|
| Interaction | Question → answer, one round trip | Give a goal → runs to completion automatically |
| Human's role | Steps in on every turn | Kicks it off, optionally approves midway |
| What model output means | The final answer shown to the user | The next action to execute (a tool call) |
| Access to the outside world | None, or read-only search | Execution rights (edit files, pay, book, etc.) |
| Blast radius of a mistake | One wrong sentence | A wrong action compounds across steps |
② The three components an agent is built from
It looks elaborate from the outside, but strip an agent down and it reduces to three parts.
1. Function calling (tool use)
When the model outputs text shaped like "call this function with these arguments," the surrounding program actually executes it and returns the result to the model. The model itself is still just doing "predict the next token" — the only thing that changed is that the tokens are now commands a program reads, not a sentence a person reads.
2. Planning — breaking a goal into steps
Given "book me the cheapest flight to Paris," an agent can't handle that in one shot. It breaks the goal down itself into search, compare, select, book. That decomposition is also just something the model writes out as text — it drafts a plan like "do A first, then decide between B and C based on the result."
3. The loop — deciding the next move from the last result
The result of a function call gets appended back into the model's input. If a search result doesn't match expectations, it retries with a different query; if it succeeds, it moves to the next step. The loop repeats until the model itself decides the goal is met, and only then does it stop. That nobody sets the stopping point in advance is the decisive difference from a chatbot.
Function calling is an old programming concept, and the plan-execute loop has been standard in robotics for decades. What made 2025–26 the moment it took off is that model judgment finally got good enough to tie the three together reliably. Earlier-generation models drafted plans too, but wandered off course mid-execution often enough to make the whole thing impractical.
③ How it actually runs — a scenario
Give an agent "plan me a trip, budget 1.5 million won," and internally it plays out roughly like this.
| Step | What the model does |
|---|---|
| 1 | Decomposes the request into a plan — flights, lodging, budget allocation, in that order |
| 2 | Calls a flight search API (a function call happens) |
| 3 | Takes the search results as input and judges whether they fit the budget |
| 4 | Calls a hotel search API with the remaining budget — the previous step's result sets this call's parameters |
| 5 | If flight + hotel together exceed the budget, it loops back to step 1 with adjusted constraints and searches again |
| 6 | Once constraints are satisfied, presents the final plan to the user (or executes the booking directly, depending on settings) |
Step 5 is the one to notice. Catching its own bad plan and backtracking — a chatbot never does that. A chatbot would just say "that's over budget" in one sentence and stop. An agent replaces that sentence with an action: a re-search.
④ Why it's suddenly everywhere now
Function calling itself has existed since 2023. Practical agents flooding in is a 2025-and-later phenomenon. Three things converging explains it accurately.
- Context expansion — as usable input length grew, a plan, search results, and intermediate judgments could all fit without the conversation breaking
- Tool-use standardization — function-calling formats that used to differ by provider converged on shared industry conventions, making it far easier for developers to wire up tools
- Better judgment — the accuracy of deciding "keep pushing the current plan, or back off" crossed the threshold into being usable in production
⑤ For investors — what agents scale up is inference demand
How an AI model gets built explained that training demand and inference demand move on different cycles. Agents push specifically on the inference side.
When a person asks a chatbot one question, that's one model call. When the same person gives an agent one goal, the plan-search-judge-retry loop from the scenario above turns that into dozens of model calls. Even with the same number of users, the inference requests generated per user grows structurally.
If AI infrastructure companies start using the phrase "agentic workloads" more often, that's a signal that inference demand can grow not just from more users but from more requests per user. How that structure translates into actual capex is covered in our circular financing analysis, and the hardware side absorbing that demand in our semiconductor sector deep dive.
⑥ The cost of autonomy — risk grows with it
A chatbot's mistake ends as one sentence, which a person reads and filters. An agent's mistake becomes an action. A bad judgment becomes the input to the next step, and another action stacks on top of it — AI hallucination is "a wrong sentence" in a chatbot, but can become "a wrong booking, a wrong file edit, a wrong payment" in an agent.
On top of that, agents carry a risk chatbots don't. The search results, documents, and web pages an agent processes can contain instructions planted by an attacker, not a person. The model has no structural way to tell those apart from the user's original instructions — this is called prompt injection, and it's the first thing to weigh when deciding how much execution authority to grant an agent. Exactly how this attack works is covered in Prompt Injection and AI Security.
The line between what to hand to an agent and what a human must check is "is this action reversible?" Search, drafting, comparative analysis — anything easily undone — can be given generous autonomy. Payment, sending, deletion — anything hard to undo — should always keep a human final-approval step, which remains the safest default for now.
A chatbot answers a question; an agent takes a goal and works through it on its own, in a loop. Function calling is the one thing that makes the loop possible; planning and looping are structure built on top of it. And as that structure scales up, what grows isn't intelligence — it's both the volume of inference requests and the blast radius when something goes wrong.
※ This article describes the general architecture of publicly documented AI agent technology as of August 2026. Individual commercial products differ in implementation. Companies mentioned are illustrative for the architecture and this is not investment advice.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.