🤖 Build an AI Agent with smolagents · สร้าง AI Agent ด้วย smolagents · 用 smolagents 做一个 AI 智能体

1 📦 Get ready — Ollama, a model, one library · เตรียมตัว — Ollama โมเดล และไลบรารีหนึ่งตัว · 准备 — Ollama、一个模型、一个库

terminal

ollama pull qwen3                  # the brain, about 5 GB. Skip it if you have it.
pip install "smolagents[litellm]"

2 🐣 Your first agent — six lines · agent ตัวแรกของคุณ — หกบรรทัด · 你的第一个智能体 — 六行

agent.py

from smolagents import CodeAgent, LiteLLMModel

model = LiteLLMModel(
    model_id="ollama_chat/qwen3",   # the model you pulled
    api_base="http://localhost:11434",    # Ollama, on your own machine
    num_ctx=8192,                         # room for the agent's long instructions
)

agent = CodeAgent(tools=[], model=model)
agent.run("How many seconds are there in one week?")

3 🔍 Read the Python it wrote · อ่านโค้ด Python ที่มันเขียน · 看看它写的 Python

save_code.py: the whole script, top to bottom

from smolagents import CodeAgent, LiteLLMModel

model = LiteLLMModel(
    model_id="ollama_chat/qwen3",
    api_base="http://localhost:11434",
    num_ctx=8192,
)
agent = CodeAgent(tools=[], model=model)

# 1) run the agent: it writes Python and runs it (watch the terminal)
agent.run("How many seconds are there in one week?")

# 2) print the code from each step
for step in agent.memory.steps:
    if getattr(step, "code_action", None):          # only steps that ran code
        print(f"--- step {step.step_number}")
        print(step.code_action)

# 3) save all of it to a NEW file, agent_code.py, in this same folder
#    "w" = write: it makes the file, or replaces an old one
code = agent.memory.return_full_code()
with open("agent_code.py", "w", encoding="utf-8") as f:
    f.write(code)
print("Saved agent_code.py")

# 4) replay the whole run, with the model's thinking
agent.replay()

terminal

python save_code.py
cat agent_code.py          # Mac and Linux: show the new file
type agent_code.py         # Windows: show the new file

what our terminal showed after the answer

Final answer: 604800
--- step 1
# Calculate seconds in a week
seconds_per_minute = 60
minutes_per_hour = 60
hours_per_day = 24
days_per_week = 7

total_seconds = days_per_week * hours_per_day * minutes_per_hour * seconds_per_minute
final_answer(total_seconds)
Saved agent_code.py

agent_code.py from our run (qwen3 on a laptop)

# Calculate seconds in a week
seconds_per_minute = 60
minutes_per_hour = 60
hours_per_day = 24
days_per_week = 7

total_seconds = days_per_week * hours_per_day * minutes_per_hour * seconds_per_minute
final_answer(total_seconds)

The model writes new code on every run, so your file will not match ours exactly. Our first run named its steps differently, and got the same 604800. โมเดลเขียนโค้ดใหม่ทุกครั้งที่รัน ไฟล์ของคุณจึงจะไม่เหมือนของเราทุกตัวอักษร การรันครั้งแรกของเราตั้งชื่อตัวแปรต่างออกไป แต่ได้ 604800 เหมือนกัน 模型每次运行都会写新的代码，所以你的文件不会和我们的一模一样。我们第一次运行时变量名起得不一样，但同样得到 604800。

4 🧰 Give it a tool · ให้เครื่องมือกับมัน · 给它一个工具

timetable_agent.py

from smolagents import CodeAgent, LiteLLMModel, tool

TIMETABLE = {
    "Monday":    ["English", "Maths", "Python", "Art"],
    "Tuesday":   ["Science", "Python", "Thai", "Python"],
    "Wednesday": ["Maths", "English", "Music"],
    "Thursday":  ["Python", "Python", "Python", "Sport"],
    "Friday":    ["Thai", "Science", "English"],
}

@tool
def get_timetable(day: str) -> list:
    """Returns the list of lessons for one school day.

    Args:
        day: A weekday name, like "Monday".
    """
    return TIMETABLE.get(day.capitalize(), [])

model = LiteLLMModel(model_id="ollama_chat/qwen3",
                     api_base="http://localhost:11434", num_ctx=8192)

agent = CodeAgent(tools=[get_timetable], model=model)
agent.run("Which day has the most Python lessons?")

5 🎭 Give it a role · ให้บทบาทกับมัน · 给它一个角色

in agent.py

agent = CodeAgent(
    tools=[],
    model=model,
    instructions=("You are Kru Nari, a friendly English teacher at a school in Chiang Mai. "
                  "Answer in short, simple English."),
)

The next section gives Kru Nari something to read, and one more rule: look it up before you answer. หัวข้อถัดไปจะให้ครูนารีมีเอกสารให้อ่าน และกฎอีกหนึ่งข้อ คือค้นก่อนตอบ 下一节给 Kru Nari 一些可以读的东西，再加一条规则：先查，再答。

6 📚 Let it read your own documents · ให้มันอ่านเอกสารของคุณเอง · 让它读你自己的文件

notes/exams.txt (one of our three example files)

English speaking test

The Grade 9 English speaking test is on Friday 19 September, in Room 204, starting at 9:00.

Bring your student card and a pencil. Phones must stay in your bag.

Each student speaks for five minutes about a photo. The teacher gives the photo on the day.

terminal

pip install rank-bm25

school_agent.py

from pathlib import Path
from rank_bm25 import BM25Okapi
from smolagents import CodeAgent, LiteLLMModel, tool

# 1) the index: every paragraph of every .txt file in notes/
chunks = []
for path in sorted(Path("notes").glob("*.txt")):
    for para in path.read_text(encoding="utf-8").split("\n\n"):
        if para.strip():
            chunks.append((path.name, para.strip()))
index = BM25Okapi([text.lower().split() for _, text in chunks])

# 2) a tool that searches the index
@tool
def search_notes(query: str) -> str:
    """Searches the school's notes and returns the three most useful
    paragraphs, each with the name of the file it came from.

    Args:
        query: What to look for, in plain words.
    """
    scores = index.get_scores(query.lower().split())
    best = sorted(range(len(chunks)), key=lambda i: scores[i], reverse=True)[:3]
    return "\n\n".join(f"[{chunks[i][0]}] {chunks[i][1]}" for i in best)

# 3) the agent: the role from section 5, plus one rule
model = LiteLLMModel(model_id="ollama_chat/qwen3",
                     api_base="http://localhost:11434", num_ctx=8192)
agent = CodeAgent(
    tools=[search_notes],
    model=model,
    instructions=("You are Kru Nari, a friendly English teacher at a school in Chiang Mai. "
                  "Answer in short, simple English. Always use search_notes before you answer "
                  "a question about the school, and say which file you used."),
)
agent.run("When is the English speaking test, and what should I bring?")

7 🔲 A tool that makes a file — QR codes · เครื่องมือที่สร้างไฟล์ — คิวอาร์โค้ด · 会生成文件的工具 — 二维码

terminal

pip install "qrcode[pil]"

qr_agent.py

import qrcode
from smolagents import CodeAgent, LiteLLMModel, tool

@tool
def make_qr_code(url: str, filename: str) -> str:
    """Makes a QR code picture for a web address and saves it as a PNG file.

    Args:
        url: The web address the QR code should open, like "https://krueng.ai/".
        filename: The name to save it under, ending in .png.
    """
    qrcode.make(url).save(filename)
    return f"Saved {filename}"

model = LiteLLMModel(model_id="ollama_chat/qwen3",
                     api_base="http://localhost:11434", num_ctx=8192)
agent = CodeAgent(tools=[make_qr_code], model=model)
agent.run("Make a QR code for https://krueng.ai/ and save it as krueng.png. "
          "Then tell me the file name.")

what the agent ran — both lines in one step (qwen3 on a laptop)

make_qr_code(url="https://krueng.ai/", filename="krueng.png")
final_answer("krueng.png")
Final answer: krueng.png

8 ⚖️ Code or JSON? Swap one word · โค้ดหรือ JSON? เปลี่ยนคำเดียว · 代码还是 JSON？换一个词

from smolagents import ToolCallingAgent

agent = ToolCallingAgent(tools=[get_timetable], model=model)
agent.run("How many lessons are there on Wednesday?")

9 🛡️ It runs code on your computer — be careful · มันรันโค้ดบนเครื่องคุณ — ต้องระวัง · 它在你电脑上运行代码 — 要小心

agent = CodeAgent(tools=[get_timetable], model=model,
                  additional_authorized_imports=["csv"])   # allow one more module

what you see if the model tries to import os

InterpreterError: Import of os is not allowed.

10 🌐 Feed it the web — the built-in search tools · ป้อนเว็บให้มัน — เครื่องมือค้นหาที่มีมาให้ · 把网络喂给它 — 自带的搜索工具

terminal

pip install "smolagents[toolkit]"     # the web tools: ddgs and markdownify

web_agent.py

from smolagents import CodeAgent, LiteLLMModel, DuckDuckGoSearchTool, VisitWebpageTool

model = LiteLLMModel(model_id="ollama_chat/qwen3",
                     api_base="http://localhost:11434",
                     num_ctx=16384)                      # web pages are long: more room

agent = CodeAgent(
    tools=[DuckDuckGoSearchTool(max_results=5),          # find pages
           VisitWebpageTool(max_output_length=5000)],    # read one, cut to 5,000 characters
    model=model,
    max_steps=6,
)

agent.run("What is the newest version of the smolagents Python library, "
          "and when was it released? Search the web, then say which page "
          "you got the answer from.")

11 👉 Now make it yours · ทำให้เป็นของคุณเอง · 把它变成你自己的

When a word here is new, go back to Start Python. When your agent grows, Catch Your Own Bugs shows you how to test it. ถ้าเจอคำที่ยังไม่รู้ กลับไปที่ เริ่มเขียน Python เมื่อ agent ของคุณใหญ่ขึ้น จับบั๊กของตัวเอง จะสอนวิธีทดสอบมัน 遇到不认识的词，就回开始学 Python。等你的智能体变大了，抓住自己的 bug 教你怎么测试它。