A Tool Is Just an Ordinary Function
Don't be scared by the word "tool." A tool you give an Agent is just a normal function you wrote, plus a line saying "what this function does," so the large model knows when to use it.
def get_weather(city: str) -> str:
"""查询某个城市今天的天气。"""
return weather_api(city)
tools = [get_weather] # 把工具交给 AgentWhat a Minimal Agent Looks Like
Write the "loop" as code and the skeleton is this small: the model decides whether to call a tool; if it does, you feed the result back; repeat until it says "I'm done answering."
while True:
step = model.think(history, tools)
if step.is_final: # 模型说：可以回答了
return step.answer
result = call_tool(step) # 做：执行工具
history.append(result) # 看：把结果喂回去，进入下一轮How Does the Model Know Which Tool to Call?
You give it a row of tools (check weather, query database, send email…), and it decides which one to use this time based on each tool's name and that one-line description—like finding a wrench in your toolbox by its label.
So the clearer the name and description, the more accurately it picks.
自测 · 学完检查一下
想真正动手做题、记进度、攒连胜？到互动课里练。
Judge: is the "tool" definition below acceptable? (Can the Agent know what it does and when to use it?)
def f(x):
return db(x)答案：Not acceptable
The name `f` and parameter `x` are unclear, and there's no description of what it does. The model decides when to call by name and description—you must give it a clear name plus a one-line description of its purpose.
What is the single most important thing a "good tool" definition must have?
答案：A clear name + a one-line doc describing what it's for
The model decides whether and when to use a tool by "name + description." Writing the description clearly matters more than anything.
In the code, what does the line `if step.is_final:` check?
答案：The model thinks it can now give the final answer
is_final = the model says "enough, I can answer," so it breaks out of the loop and returns the answer.
In the minimal Agent loop, after calling a tool and getting a result, you have to ___ the result back to the model so it can move to the next round of thinking.
答案：feed
Add the tool result back to the conversation history so the model can see the result of the "act"—that's what keeps the loop turning.
Given a row of tools, what does the Agent rely on to decide which one to call this time?
答案：Each tool's name and description
Name + description is the model's "manual." That's also why `f(x)` in question 1 was unacceptable.
Judge: this tool list has clear names and descriptions—is it acceptable?
def get_order_count(day: str) -> int:
"""查询某天的总下单量。"""
...
def send_email(to: str, body: str) -> None:
"""给指定邮箱发一封邮件。"""
...答案：Acceptable
The names are self-explanatory and each has a one-line purpose—the model can accurately judge when to use which. That's an acceptable tool definition.