03 · The AI agent loop

AI agent loop video

Every AI agent runs on one tiny loop: think, act, observe, repeat. Finance asks “what is 15% of Q3 revenue?”. The model alone can’t know that number, so we give it tools, let it decide what to do at each step, and write what it learns back into memory. Three steps later it answers: $630,000.

Watch: YouTube · TikTok · Instagram (83 seconds)

The idea

An agent is a language model in a loop:

Think: the model reads everything in memory and picks an action, like “search for Q3 revenue”.
Act: your code runs the tool it asked for.
Observe: the result goes back into memory, so the next thought knows more.

The loop stops when the model decides it can answer, or when it hits a step limit.

Run it

python3 src/agent.py

0 search -> Q3 revenue: 4200000
1 calc -> 630000.0
answer: $630,000

Files

03-ai-agent-loop/
├── data/
│   └── warehouse.json    the numbers the search tool can find
└── src/
    ├── agent.py          the loop from the video
    ├── model.py          think(): stand-in for the LLM
    └── tools.py          search and calculator

agent.py is the loop, and the file you see in the video.
tools.py has two tools. search looks a number up in warehouse.json, and calculator multiplies.
model.py has think(), a stand-in for the LLM. It follows simple rules so the example runs offline and gives the same answer every time. In a real agent, think() is one API call to a model that returns the same (action, argument) shape.

The code

Line	Code	What it does
1	`from tools import search, calculator`	The agent’s tools. In a company these are real APIs.
2	`from model import think`	The “brain”. Here a stand-in, in production an LLM call.
4	`goal = "What is 15% of Q3 revenue?"`	The question from finance.
5	`memory = [goal]`	Memory starts with just the goal.
7	`for step in range(5):`	At most 5 steps. This is the guardrail (more below).
8	`action, arg = think(memory)`	Think: read memory, pick the next action.
9 to 11	`if action == "answer": ... break`	If the model is ready, print the answer and stop.
12 to 13	`tool = {...}[action]`	Otherwise look up the tool it asked for.
14	`result = tool(arg)`	Act: run the tool.
15	`memory.append(result)`	Observe: write the result into memory.
16	`print(step, action, "->", result)`	Show each step.

Why line 7 matters

range(5) is a guardrail. A real model can get confused and keep calling tools forever, burning time and money. Production agents always have limits: on steps, on tokens, on cost, and on which tools they may call.

Try this

Ask for Q2 instead: change the goal and update model.py so the search uses “Q2 revenue”.
Add a third tool, for example format_currency, and make think() use it before answering.
Break the guardrail: make think() never return "answer". What happens with and without range(5)?
Replace think() with a real LLM call that returns JSON like {"action": "search", "arg": "Q3 revenue"}.

Previous: 02 · K-means · Next: 04 · Ontologies · All lessons