LangChain自定义ReAct代理无限循环问题求助（报错：Missing 'Action:' after 'Thought:'）

我看了你的代码和终端报错，这个无限循环+格式错误的问题，核心是自定义的ReAct Prompt模板没对齐LangChain ReAct Agent的格式要求，再加上几个参数传递的小失误导致的。下面给你一步步拆解修复：

问题1：Prompt模板格式不符合LangChain的解析规则 #

LangChain的ReAct Agent对LLM的输出格式有严格约定，必须明确输出Action: 工具名和Action Input: 参数这样的标准标记。你的模板里的示例写法（比如[Action]: I should use the get_weather tool...）太口语化，解析器根本识别不出合规的Action标记，所以才会一直报错“Missing 'Action:' after 'Thought:'”，进而陷入循环。

修复后的Prompt模板：
要明确告诉LLM必须遵循的格式，把示例改成标准写法，同时简化思考流程的描述：

react_template = """
Answer the following questions as best you can, using the tools provided.

Tools:
{tools}

Use the following format:

Question: the input question you must answer
Thought: Do I need to use a tool? Yes/No
Action: the tool to use, should be one of [{tool_names}]
Action Input: the input to give the tool
Observation: the result of the tool
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin!

Question: {question}
{agent_scratchpad}
"""

问题2：Prompt的input_variables遗漏了agent_scratchpad #

你定义的react_prompt里，input_variables只包含了question、tools、tool_names，但agent_scratchpad是Agent用来记录历史思考/操作的关键变量，没加的话Agent无法迭代更新状态，也是循环的诱因之一。

修复：

react_prompt = PromptTemplate(
    template=react_template,
    input_variables=["question", "tools", "tool_names", "agent_scratchpad"],  # 补上agent_scratchpad
)

问题3：调用AgentExecutor时手动传入了多余参数 #

你在invoke的时候手动传了tools、tool_names、agent_scratchpad，但这些参数已经由create_react_agent自动绑定到Prompt里了，手动传入会覆盖正确的工具信息，导致Agent无法正确识别可用工具。

修复：
invoke只需要传question即可：

if __name__ == "__main__":
    response = agent_exexcutor.invoke(
        input={"question": "What is the weather in france ?"}
    )
    print(response["output"])

可选优化：降低LLM温度减少随机性 #

你用的是llama3.2:1b小模型，即使temperature=0.1，偶尔还是会输出不符合格式的内容，可以把温度调到0.0，进一步约束输出的一致性：

llm = ChatOllama(
    model="llama3.2:1b",
    temperature=0.0,  # 改成0.0
)

完整修复后的代码 #

把上面的修改整合后，完整代码如下：

from langchain_ollama import ChatOllama
from langchain.tools import Tool
from langchain.agents import create_react_agent, AgentExecutor
from langchain_core.prompts import PromptTemplate

llm = ChatOllama(
    model="llama3.2:1b",
    temperature=0.0,
)

react_template = """
Answer the following questions as best you can, using the tools provided.

Tools:
{tools}

Use the following format:

Question: the input question you must answer
Thought: Do I need to use a tool? Yes/No
Action: the tool to use, should be one of [{tool_names}]
Action Input: the input to give the tool
Observation: the result of the tool
... (this Thought/Action/Action Input/Observation can repeat N times)
Thought: I now know the final answer
Final Answer: the final answer to the original input question

Begin!

Question: {question}
{agent_scratchpad}
"""

react_prompt = PromptTemplate(
    template=react_template,
    input_variables=["question", "tools", "tool_names", "agent_scratchpad"],
)


def get_weather_func(location: str) -> str:
    location = location.lower().strip()
    if location == "france":
        return "cloudy"
    elif location == "italy":
        return "sunny"
    else:
        return "unknown weather"

weather_tool = Tool(
    name= "get_weather",
    func=get_weather_func,
    description="Useful for getting the weather of a location"
)

tools = [weather_tool]

agent = create_react_agent(
    llm=llm,
    tools=tools,
    prompt=react_prompt
)

agent_exexcutor = AgentExecutor(
    agent=agent,
    tools=tools,
    verbose=True,
    handle_parsing_errors=True
)

if __name__ == "__main__":
    response = agent_exexcutor.invoke(
        input={"question": "What is the weather in france ?"}
    )
    
    print(response["output"])

为什么这样能解决问题？ #

调整后的Prompt严格遵循了LangChain ReAct Agent的格式规范，解析器能准确识别Action和Action Input标记，拿到工具返回的结果后，LLM会按照模板判断是否已经得到最终答案，从而跳出循环返回结果，不会再出现格式错误的情况。

备注：内容来源于stack exchange，提问作者Lakshdeep Singh