Library vs framework

When you will meet it

While choosing AI tools you keep meeting the argument between 'use a framework like LangChain' and 'call the API directly'. Without the distinction you cannot judge how much control a tool actually leaves you — and control is the dividing line between being able to trace a failure and not.

Minimal example

函式庫（你主導）：
  import requests
  r = requests.post(url, json=payload)     # 你決定什麼時候發、發什麼
  text = r.json()["choices"][0]["message"]["content"]
  # 之後每一步也都是你寫的

框架（它主導）：
  agent = create_react_agent(llm, tools, prompt)
  agent.invoke({"messages": [...]})        # 之後的迴圈由框架跑
  # 它自己決定何時呼叫模型、何時呼叫工具、何時停

The test is 'who owns the loop'. If your own code holds a while loop that calls the model, you are using libraries. If you call invoke once and something keeps spinning inside, you are in a framework. LangChain and LlamaIndex are the latter; client packages like requests or openai are the former.

What people get wrong

• Assuming a framework is 'more advanced' and therefore right for beginners. The opposite: a framework hides detail, so when it fails you cannot even see which layer broke. Get the same thing working once with libraries and you will know exactly what the framework was doing for you.
• Using 'framework' and an agent's 'harness' as the same word. A framework (like LangChain) is a parts box for building a harness; the harness is the running thing you built, which assembles context and executes tools.