> https://github.com/hongyingyue/LLM-agent-tutorials

强化学习中的智能体 通过与环境交互，学习新的策略从而获得奖励。

最近流行的大模型智能体，以大模型为大脑，通过调用工具来完成复杂的任务。

大模型工具调用

• https://github.com/hongyingyue/LLM-agent-tutorials/blob/main/1_llm_tool_call.ipynb

最简单的调用工具的例子

# Tool 1: DuckDuckGo Search
def duckduckgo_search(query, max_results=5):
    results = []
    with DDGS() as ddgs:
        for r in ddgs.text(query, max_results=max_results):
            results.append(f"{r['title']}: {r['href']}")
    return "\n".join(results)

工具注册给大模型后，根据任务，大模型会返回的是工具调用

🔧 Tool requested: duckduckgo_search({'query': 'Beijing weather forecast tomorrow'})

然后我们的agent会负责执行这个工具，并把工具结果作为context继续输入给大模型，如果需要多次调用工具则进行反复调用，最终大模型输出正式的结论。

ReACT

• https://github.com/hongyingyue/LLM-agent-tutorials/blob/main/6_react.ipynb

观察 - 思考 - 行动的范式

行动后的结果进一步加入到context从而继续观察。在prompt中，要求大模型按照特定格式进行回应:

REACT_PROMPT = """
You are a smart physics assistant.
To solve a question, you must iterate through these steps:

1. Question: The input question you must answer
2. Thought: Comment on what you want to do next
3. Action: The function to call. Available tools: [get_planet_mass, calculate_gravity]
4. Action Input: The exact arguments to the function
5. Observation: The output of the tool (I will provide this)

... (Repeat steps 2-5 as needed) ...

6. Final Answer: The answer to the original question.

Rules:
- You MUST output 'Thought:', 'Action:', and 'Action Input:' in that order.
- Stop generating after 'Action Input'. I will give you the 'Observation'.
- When you have the answer, output 'Final Answer:'.

Current conversation:
{history}
"""

记忆

• https://github.com/hongyingyue/LLM-agent-tutorials/blob/main/7_memory.ipynb
• https://github.com/mem0ai/mem0

可以分为短期记忆与长期记忆，同时Agentic RAG也可以认为是记忆模块的一种，claude的skills，Manus的文件也都是为了增强智能体的记忆能力。

上下文工程

从 Prompt Engineering 到 Context Engineering，更有针对性的管理上下文，例如retreval的内容。

# 1. 动态组装上下文
context = ContextEngine()
context.add_system_instructions(project_patterns)
context.add_retrieved_docs(relevant_components)
context.add_conversation_history(recent_messages, max_tokens=4000)
context.add_user_profile(user_preferences)
context.add_available_tools(filtered_tools)

MCP协议

• https://github.com/hongyingyue/LLM-agent-tutorials/blob/main/8_mcp_agent.ipynb

Agentic-RL

• https://github.com/hongyingyue/LLM-agent-tutorials/blob/main/11_rl_finetune.ipynb

相比RLHF, Agentic RL关注的是多轮任务决策

评估