Agent Loop：AI Agent 的核心引擎【nanobot架构分析系列】

玄同 765

大语言模型 (LLM) 开发工程师 | 中国传媒大学 · 数字媒体技术（智能交互与游戏设计）

优快云 · 个人主页 | GitHub · Follow

---

关于作者

• 深耕领域：大语言模型开发 / RAG 知识库 / AI Agent 落地 / 模型微调
• 技术栈：Python | RAG (LangChain / Dify + Milvus) | FastAPI + Docker
• 工程能力：专注模型工程化部署、知识库构建与优化，擅长全流程解决方案

「让 AI 交互更智能，让技术落地更高效」
欢迎技术探讨与项目合作，解锁大模型与智能交互的无限可能！

---

Agent Loop：AI Agent 的核心引擎

Agent Loop 是 nanobot 的心脏，负责消息处理、上下文构建、LLM 交互和工具调用的完整循环。

概述

Agent Loop 是 nanobot 的核心处理引擎，实现了"感知-思考-行动"的完整循环。它从消息总线接收用户消息，构建包含历史、记忆、技能的上下文，调用 LLM 进行推理，执行工具调用，最后将响应返回给用户。本文将深入剖析 Agent Loop 的设计与实现。

问题背景

Agent Loop 的核心挑战

构建一个 AI Agent 的核心循环需要解决以下问题：

挑战	描述	nanobot 的解决方案
消息处理	如何高效处理异步消息	消息总线 + 异步队列
上下文构建	如何组装完整的对话上下文	ContextBuilder 模块化构建
工具调用	如何处理多轮工具调用	迭代循环 + 结果反馈
会话管理	如何维护对话历史	SessionManager 持久化
错误处理	如何优雅处理异常	try-catch + 错误响应

Agent Loop 的设计目标

核心架构

AgentLoop 类结构

AgentLoop 是整个系统的核心协调者：

class AgentLoop:
    """
    The agent loop is the core processing engine.
  
    It:
    1. Receives messages from the bus
    2. Builds context with history, memory, skills
    3. Calls the LLM
    4. Executes tool calls
    5. Sends responses back
    """
  
    def __init__(
        self,
        bus: MessageBus,
        provider: LLMProvider,
        workspace: Path,
        model: str | None = None,
        max_iterations: int = 20,
        brave_api_key: str | None = None,
        exec_config: "ExecToolConfig | None" = None,
        cron_service: "CronService | None" = None,
        restrict_to_workspace: bool = False,
        session_manager: SessionManager | None = None,
    ):
        # ... 初始化代码

依赖组件

消息处理流程

主循环实现

Agent Loop 的核心是一个异步消息处理循环：

async def run(self) -> None:
    """Run the agent loop, processing messages from the bus."""
    self._running = True
    logger.info("Agent loop started")
  
    while self._running:
        try:
            # 等待下一条消息
            msg = await asyncio.wait_for(
                self.bus.consume_inbound(),
                timeout=1.0
            )
          
            # 处理消息
            try:
                response = await self._process_message(msg)
                if response:
                    await self.bus.publish_outbound(response)
            except Exception as e:
                logger.error(f"Error processing message: {e}")
                # 发送错误响应
                await self.bus.publish_outbound(OutboundMessage(
                    channel=msg.channel,
                    chat_id=msg.chat_id,
                    content=f"Sorry, I encountered an error: {str(e)}"
                ))
        except asyncio.TimeoutError:
            continue

消息处理详解

上下文构建

ContextBuilder 的职责

ContextBuilder 负责组装完整的对话上下文：

class ContextBuilder:
    """
    Builds the context (system prompt + messages) for the agent.
  
    Assembles bootstrap files, memory, skills, and conversation history
    into a coherent prompt for the LLM.
    """
  
    BOOTSTRAP_FILES = ["AGENTS.md", "SOUL.md", "USER.md", "TOOLS.md", "IDENTITY.md"]

上下文构建流程

System Prompt 构建

def build_system_prompt(self, skill_names: list[str] | None = None) -> str:
    """Build the system prompt from bootstrap files, memory, and skills."""
    parts = []
  
    # 核心身份
    parts.append(self._get_identity())
  
    # Bootstrap 文件
    bootstrap = self._load_bootstrap_files()
    if bootstrap:
        parts.append(bootstrap)
  
    # 记忆上下文
    memory = self.memory.get_memory_context()
    if memory:
        parts.append(f"# Memory\n\n{memory}")
  
    # 技能 - 渐进式加载
    # 1. Always-loaded 技能: 包含完整内容
    always_skills = self.skills.get_always_skills()
    if always_skills:
        always_content = self.skills.load_skills_for_context(always_skills)
        if always_content:
            parts.append(f"# Active Skills\n\n{always_content}")
  
    # 2. 可用技能: 只显示摘要
    skills_summary = self.skills.build_skills_summary()
    if skills_summary:
        parts.append(f"""# Skills

The following skills extend your capabilities. To use a skill, read its SKILL.md file using the read_file tool.
Skills with available="false" need dependencies installed first.

{skills_summary}""")
  
    return "\n\n---\n\n".join(parts)

核心身份信息

def _get_identity(self) -> str:
    """Get the core identity section."""
    now = datetime.now().strftime("%Y-%m-%d %H:%M (%A)")
    tz = _time.strftime("%Z") or "UTC"
    workspace_path = str(self.workspace.expanduser().resolve())
    system = platform.system()
    runtime = f"{'macOS' if system == 'Darwin' else system} {platform.machine()}, Python {platform.python_version()}"
  
    return f"""# nanobot

You are nanobot, a helpful AI assistant. You have access to tools that allow you to:
- Read, write, and edit files
- Execute shell commands
- Search the web and fetch web pages
- Send messages to users on chat channels
- Spawn subagents for complex background tasks

## Current Time
{now} ({tz})

## Runtime
{runtime}

## Workspace
Your workspace is at: {workspace_path}
- Memory files: {workspace_path}/memory/MEMORY.md
- Daily notes: {workspace_path}/memory/YYYY-MM-DD.md
- Custom skills: {workspace_path}/skills/{{skill-name}}/SKILL.md
"""

工具调用循环

迭代处理机制

Agent Loop 使用迭代循环处理多轮工具调用：

async def _process_message(self, msg: InboundMessage) -> OutboundMessage | None:
    """Process a single inbound message."""
  
    # 获取或创建会话
    session = self.sessions.get_or_create(msg.session_key)
  
    # 构建初始消息
    messages = self.context.build_messages(
        history=session.get_history(),
        current_message=msg.content,
        media=msg.media if msg.media else None,
        channel=msg.channel,
        chat_id=msg.chat_id,
    )
  
    # Agent 循环
    iteration = 0
    final_content = None
  
    while iteration < self.max_iterations:
        iteration += 1
      
        # 调用 LLM
        response = await self.provider.chat(
            messages=messages,
            tools=self.tools.get_definitions(),
            model=self.model
        )
      
        # 处理工具调用
        if response.has_tool_calls:
            # 添加助手消息（包含工具调用）
            tool_call_dicts = [
                {
                    "id": tc.id,
                    "type": "function",
                    "function": {
                        "name": tc.name,
                        "arguments": json.dumps(tc.arguments)
                    }
                }
                for tc in response.tool_calls
            ]
            messages = self.context.add_assistant_message(
                messages, response.content, tool_call_dicts,
                reasoning_content=response.reasoning_content,
            )
          
            # 执行工具
            for tool_call in response.tool_calls:
                args_str = json.dumps(tool_call.arguments, ensure_ascii=False)
                logger.info(f"Tool call: {tool_call.name}({args_str[:200]})")
                result = await self.tools.execute(tool_call.name, tool_call.arguments)
                messages = self.context.add_tool_result(
                    messages, tool_call.id, tool_call.name, result
                )
            # 交错思维链: 在下一步行动前反思
            messages.append({"role": "user", "content": "Reflect on the results and decide next steps."})
        else:
            # 无工具调用，完成处理
            final_content = response.content
            break
  
    if final_content is None:
        final_content = "I've completed processing but have no response to give."
  
    # 保存到会话
    session.add_message("user", msg.content)
    session.add_message("assistant", final_content)
    self.sessions.save(session)
  
    return OutboundMessage(
        channel=msg.channel,
        chat_id=msg.chat_id,
        content=final_content,
        metadata=msg.metadata or {},
    )

工具调用流程图

会话管理

Session 数据结构

@dataclass
class Session:
    """
    A conversation session.
  
    Stores messages in JSONL format for easy reading and persistence.
    """
  
    key: str  # channel:chat_id
    messages: list[dict[str, Any]] = field(default_factory=list)
    created_at: datetime = field(default_factory=datetime.now)
    updated_at: datetime = field(default_factory=datetime.now)
    metadata: dict[str, Any] = field(default_factory=dict)
  
    def add_message(self, role: str, content: str, **kwargs: Any) -> None:
        """Add a message to the session."""
        msg = {
            "role": role,
            "content": content,
            "timestamp": datetime.now().isoformat(),
            **kwargs
        }
        self.messages.append(msg)
        self.updated_at = datetime.now()
  
    def get_history(self, max_messages: int = 50) -> list[dict[str, Any]]:
        """Get message history for LLM context."""
        recent = self.messages[-max_messages:] if len(self.messages) > max_messages else self.messages
        return [{"role": m["role"], "content": m["content"]} for m in recent]

SessionManager 实现

class SessionManager:
    """Manages conversation sessions."""
  
    def __init__(self, workspace: Path):
        self.workspace = workspace
        self.sessions_dir = ensure_dir(Path.home() / ".nanobot" / "sessions")
        self._cache: dict[str, Session] = {}
  
    def get_or_create(self, key: str) -> Session:
        """Get an existing session or create a new one."""
        # 检查缓存
        if key in self._cache:
            return self._cache[key]
      
        # 尝试从磁盘加载
        session = self._load(key)
        if session is None:
            session = Session(key=key)
      
        self._cache[key] = session
        return session

系统消息处理

子代理通知机制

Agent Loop 还处理系统消息，如子代理完成通知：

async def _process_system_message(self, msg: InboundMessage) -> OutboundMessage | None:
    """
    Process a system message (e.g., subagent announce).
  
    The chat_id field contains "original_channel:original_chat_id" to route
    the response back to the correct destination.
    """
    # 解析来源
    if ":" in msg.chat_id:
        parts = msg.chat_id.split(":", 1)
        origin_channel = parts[0]
        origin_chat_id = parts[1]
    else:
        origin_channel = "cli"
        origin_chat_id = msg.chat_id
  
    # 使用原始会话的上下文
    session_key = f"{origin_channel}:{origin_chat_id}"
    session = self.sessions.get_or_create(session_key)
  
    # ... 处理逻辑与普通消息类似

关键设计决策

1. 最大迭代限制

max_iterations: int = 20

防止无限循环，确保系统稳定性。

2. 交错思维链

messages.append({"role": "user", "content": "Reflect on the results and decide next steps."})

在工具执行后添加反思提示，帮助 LLM 更好地规划下一步行动。

3. 错误恢复

except Exception as e:
    logger.error(f"Error processing message: {e}")
    await self.bus.publish_outbound(OutboundMessage(
        channel=msg.channel,
        chat_id=msg.chat_id,
        content=f"Sorry, I encountered an error: {str(e)}"
    ))

确保错误不会导致系统崩溃，并向用户提供有意义的反馈。

性能优化

1. 会话缓存

self._cache: dict[str, Session] = {}

避免频繁的磁盘 I/O，提高响应速度。

2. 历史消息限制

def get_history(self, max_messages: int = 50) -> list[dict[str, Any]]:

限制上下文长度，避免超出 LLM 的 token 限制。

3. 异步处理

async def run(self) -> None:
    while self._running:
        msg = await asyncio.wait_for(
            self.bus.consume_inbound(),
            timeout=1.0
        )

使用异步 I/O 提高并发处理能力。

总结

Agent Loop 是 nanobot 的核心引擎，其设计体现了以下原则：

设计原则	实现方式
模块化	ContextBuilder、SessionManager、ToolRegistry 各司其职
可扩展	通过注册机制动态添加工具和技能
可靠性	迭代限制、错误恢复、会话持久化
可观测	完善的日志记录和状态追踪

Agent Loop 的核心是一个简洁但强大的迭代循环，通过上下文构建、LLM 调用和工具执行的组合，实现了完整的 Agent 功能。

---