Dify 在 core/agent/entities.py 里把策略定成一个枚举（core/agent/entities.py:73）：

# api/core/agent/entities.py:68
class AgentEntity(BaseModel):
    class Strategy(StrEnum):
        CHAIN_OF_THOUGHT = "chain-of-thought"     # ← CoT：靠提示词约定的文本格式
        FUNCTION_CALLING = "function-calling"      # ← FC：靠模型原生的工具调用能力

    provider: str
    model: str
    strategy: Strategy                             # 用哪种策略
    prompt: AgentPromptEntity | None = None
    tools: list[AgentToolEntity] | None = None     # 这个 Agent 能用哪些工具
    max_iteration: int = 10                        # ★最多循环几轮（防跑飞）

两种 Runner 都继承 BaseAgentRunner（core/agent/base_agent_runner.py:51）。底座把"和策略无关的公共活"都干了。构造函数（base_agent_runner.py:52）接收一堆运行环境：

# api/core/agent/base_agent_runner.py:52
def __init__(self, *, session, tenant_id,
             application_generate_entity: AgentChatAppGenerateEntity,
             conversation: Conversation,
             app_config: AgentChatAppConfig,
             model_config: ModelConfigWithCredentialsEntity, ...):
    ...   # 存下 租户/会话/应用配置/模型配置，供两种策略共用

最关键的公共方法是"把工具翻译给模型看"——_init_prompt_tools（base_agent_runner.py:188）和它内部的 _convert_tool_to_prompt_message_tool（base_agent_runner.py:138）：

看 CotAgentRunner.run（core/agent/cot_agent_runner.py:49）开头，两个细节最能说明 CoT 的精髓：

# api/core/agent/cot_agent_runner.py:49
def run(self, session, message, query, inputs):
    ...
    # check model mode
    if "Observation" not in app_generate_entity.model_conf.stop:      # ★① 把 "Observation" 设成停止词
        if ...provider not in self._ignore_observation_providers:
            app_generate_entity.model_conf.stop.append("Observation")

    iteration_step = 1
    max_iteration_steps = min(app_config.agent.max_iteration, 99) + 1  # ★② 最多循环几轮（封顶 99）
    tool_instances, prompt_messages_tools = self._init_prompt_tools()  #    准备工具菜单
    function_call_state = True
    ...

而模型输出的"行动"被解析成一个 Action（cot_agent_runner.py:36），循环里据它决定：是"最终答案"就结束，否则调工具（cot_agent_runner.py:227 附近）：

# api/core/agent/cot_agent_runner.py (循环体节选)
if scratchpad.action.action_name.lower() == "final answer":
    ...                                        # ★ 模型说"最终答案" → 收尾返回
else:
    function_call_state = True                 # ★ 否则是要调工具 → 继续循环
    tool_invoke_response, tool_invoke_meta = self._handle_invoke_action(
        session=session, action=scratchpad.action,
        tool_instances=tool_instances, ...)
    scratchpad.observation = tool_invoke_response   # 把工具结果记成"观察"
    self.save_agent_thought(...)                    # 存这一轮的思考+行动+观察

FunctionCallAgentRunner.run（core/agent/fc_agent_runner.py:37）不用约定文本格式，而是读模型返回里结构化的 tool_calls。它的主循环骨架（fc_agent_runner.py:81）：

# api/core/agent/fc_agent_runner.py:55
iteration_step = 1
max_iteration_steps = min(app_config.agent.max_iteration, 99) + 1
function_call_state = True

while function_call_state and iteration_step <= max_iteration_steps:   # ★ 循环，带上限
    function_call_state = False
    if iteration_step == max_iteration_steps:
        prompt_messages_tools = []          # 最后一轮：把工具收走，逼模型直接给答案
    ...
    if self.check_tool_calls(chunk):        # ★ 模型这轮吐了 tool_calls 吗？
        function_call_state = True          #   吐了 → 还要再循环
        tool_calls.extend(self.extract_tool_calls(chunk) or [])   # 抽出(id, 名字, 参数)

抽出工具调用后，逐个真的执行（fc_agent_runner.py:234）——注意它直接接上了 Day16 的 ToolEngine.agent_invoke：

# api/core/agent/fc_agent_runner.py:234
for tool_call_id, tool_call_name, tool_call_args in tool_calls:
    tool_instance = tool_instances.get(tool_call_name)
    if not tool_instance:
        tool_response = {... "tool_response": f"there is not a tool named {tool_call_name}"}  # 模型点了不存在的工具
    else:
        tool_invoke_response, message_files, tool_invoke_meta = ToolEngine.agent_invoke(  # ★ 真调工具
            session=session, tool=tool_instance, tool_parameters=tool_call_args,
            user_id=self.user_id, tenant_id=self.tenant_id, message=self.message, ...)
    tool_responses.append(tool_response)     # 结果收集起来，喂回模型继续下一轮

两种 Runner 开头都有同一句 max_iteration_steps = min(app_config.agent.max_iteration, 99) + 1。这是 Agent 最重要的"保险丝"。FC 版还在到顶时直接报错（fc_agent_runner.py:229）：

# api/core/agent/fc_agent_runner.py:228 附近
# Check if max iteration is reached and model still wants to call tools
if iteration_step == max_iteration_steps and tool_calls:
    raise AgentMaxIterationError(app_config.agent.max_iteration)   # ★ 到顶还想调工具 → 报错止损