What is an AI agent?

By the end of this page — you can tell a chatbot from an agent, name the three steps of the agent loop, and explain why the loop is both the superpower and the risk.

An LLM by itself is a one-shot machine: you send text, it sends text back. Ask it “what’s on my disk?” and it can only guess — it has no disk, no eyes, no hands.

An agent is the same model wired into a loop with tools:

        ┌──────────────────────────────┐
        │                              ▼
   ┌─────────┐   tool call    ┌─────────────┐   result   ┌───────────┐
   │ THINK   │ ─────────────▶ │ ACT (tool)  │ ──────────▶ │ OBSERVE   │
   │ plan,   │                │ read file,  │             │ read the  │
   │ decide  │                │ run command │             │ outcome   │
   └─────────┘                └─────────────┘             └─────┬─────┘
        ▲                                                        │
        └────────────────────────────────────────────────────────┘

Step	The model does	Example
1. Think	Reasons about the goal, picks the next move	“I don’t know the file list — let me check”
2. Act	Emits a structured tool call (JSON)	`list_files("src/")`
3. Observe	Reads the tool’s result, added to its context	`"main.py, utils.py, test_main.py"`
4. Decide	Loop again, or stop and answer	“Found it — now let me read main.py”

Why the loop is the superpower

Multi-step jobs. “Summarize this repo” needs: list files → read the big ones → draft → polish. No single chat turn can do that; a loop can.
Self-correction. The tool returns an error? The model sees the error, reads it, and retries differently. A one-shot chat would have just hallucinated an answer.
Growing context. Each observation adds real, verified facts to the conversation — the agent learns from its own actions (within the context window).

Why the loop is also the risk

Errors compound — Every step is a small probability of a mistake. Ten steps at 95% accuracy end up at **~60%** — the whole job succeeds only if *every* step succeeds. Longer loops = more power **and** more places to go wrong.

The kitchen analogy — The model is the chef's brain. The tools are knives, stove and oven. The loop is the taste-and-adjust cycle. And the **harness** is the kitchen itself — where all of it is allowed to happen.

Next stop: what that kitchen — the harness — actually does for you.