如何监控AI代理工作流？2026年LangChain与CrewAI集成指南

引言

在现代人工智能应用开发中，尤其是在构建复杂的多代理系统时，可观测性、审计追踪记录智能体工作流中所有操作和决策的完整历史，用于调试、分析和合规性检查。和合规性报告已成为不可或缺的需求。开发者需要清晰地了解代理的决策过程、工具调用链以及任务执行的生命周期。ContextGraph Cloud一个AI智能体监控和日志记录平台，用于追踪和分析智能体工作流的执行过程。正是为解决这些挑战而设计，而 contextgraph-integrations 项目则提供了官方 Python 集成，让开发者能够轻松地将这些能力注入到主流的 AI 框架中。

在现代人工智能应用开发中，尤其是在构建复杂的多代理系统时，可观测性、审计追踪记录智能体工作流中所有操作和决策的完整历史，用于调试、分析和合规性检查。和合规性报告已成为不可或缺的需求。开发者需要清晰地了解代理的决策过程、工具调用链以及任务执行的生命周期。ContextGraph Cloud一个AI智能体监控和日志记录平台，用于追踪和分析智能体工作流的执行过程。正是为解决这些挑战而设计，而 contextgraph-integrations 项目则提供了官方 Python 集成，让开发者能够轻松地将这些能力注入到主流的 AI 框架中。

可用集成

contextgraph-integrations 项目目前为两个主流的 AI 代理框架提供了官方支持包：

包名	框架	安装命令
`contextgraph-langchain`	LangChainA framework for developing applications powered by language models through composable components.	`pip install contextgraph-langchain`
`contextgraph-crewai`	CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。	`pip install contextgraph-crewai`

The contextgraph-integrations project currently provides official support packages for two mainstream AI agent frameworks:

Package	Framework	Install Command
`contextgraph-langchain`	LangChainA framework for developing applications powered by language models through composable components.	`pip install contextgraph-langchain`
`contextgraph-crewai`	CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。	`pip install contextgraph-crewai`

快速开始

LangChainA framework for developing applications powered by language models through composable components. 集成

对于使用 LangChainA framework for developing applications powered by language models through composable components.（特别是较新版本）的开发者，可以通过 ContextGraphCallback 回调处理器在LangChain中用于拦截和处理框架事件的组件，用于记录智能体执行过程。轻松集成。项目已更新至支持 LangChainA framework for developing applications powered by language models through composable components. v1 及以上的中间件在LangChain v1+中引入的架构模式，用于在模型调用前后插入自定义逻辑。模式，同时也兼容旧版的回调模式。

For developers using LangChainA framework for developing applications powered by language models through composable components. (especially newer versions), integration can be easily achieved through the ContextGraphCallback callback handler. The project has been updated to support the middleware pattern of LangChainA framework for developing applications powered by language models through composable components. v1 and above, while also maintaining compatibility with the older callback pattern.

from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent
from contextgraph_langchain import ContextGraphCallback

# 初始化回调处理器
callback = ContextGraphCallback(
    api_key="your-api-key",  # 您的 ContextGraph API 密钥
    agent_id="my-agent"      # 用于标识此代理的唯一 ID
)

# 构建您的 LLM 和代理
llm = ChatOpenAI(model="gpt-4o")
agent = create_react_agent(llm, tools) # `tools` 是您定义的工具列表

# 运行代理并附加回调
result = agent.invoke(
    {"messages": [("user", "What's the weather?")]},
    config={"callbacks": [callback]}
)

CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。集成

对于基于 CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。框架构建的多代理协作系统，可以通过 ContextGraphObserver 观察者进行集成，从而追踪整个 Crew 中所有代理和任务的执行情况。

For multi-agent collaborative systems built on the CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。 framework, integration can be done via the ContextGraphObserver, enabling tracking of the execution of all agents and tasks within the entire Crew.

from crewai import Crew
from contextgraph_observer import ContextGraphObserver

# 初始化观察者
observer = ContextGraphObserver(
    api_key="your-api-key",  # 您的 ContextGraph API 密钥
    crew_id="my-crew"        # 用于标识此 Crew 的唯一 ID
)

# 构建您的 Crew
crew = Crew(
    agents=[agent1, agent2],
    tasks=[task1, task2],
    callbacks=[observer]     # 将观察者添加到回调列表
)

# 启动任务
result = crew.kickoff()

记录内容

每个集成都会自动捕获以下关键信息，为您的 AI 工作流提供完整的上下文：

工具调用 - 调用了哪些工具以及调用原因。
工具执行 - 输入参数、输出结果以及执行过程中出现的错误。
代理推理 - 代理的思考过程和决策逻辑。
任务生命周期 - 任务的开始、完成、失败等状态事件。

Each integration automatically captures the following key information, providing complete context for your AI workflows:

Tool Invocations - What tools are being called and why.
Tool Executions - Input parameters, output results, and any errors that occur during execution.
Agent Reasoning - The agent's thought process and decision logic.
Task Lifecycle - Status events such as task start, completion, and failure.

所有事件都会被记录并发送到 ContextGraph Cloud一个AI智能体监控和日志记录平台，用于追踪和分析智能体工作流的执行过程。平台，形成完整的上下文链条，主要用于：

审计追踪记录智能体工作流中所有操作和决策的完整历史，用于调试、分析和合规性检查。 - 满足内部审计或监管合规要求。
策略执行 - 监控和确保代理行为符合预设策略。
合规性报告 - 生成用于演示合规性的详细报告。
调试与可观测性 - 深入理解系统行为，快速定位和解决问题。

All events are logged and sent to the ContextGraph Cloud一个AI智能体监控和日志记录平台，用于追踪和分析智能体工作流的执行过程。 platform, forming a complete contextual chain, primarily used for:

Audit Trails - Meeting internal audit or regulatory compliance requirements.
Policy Enforcement - Monitoring and ensuring agent behavior aligns with predefined policies.
Compliance Reporting - Generating detailed reports to demonstrate compliance.
Debugging and Observability - Gaining deep insights into system behavior for rapid issue identification and resolution.

获取 API 密钥

要开始使用 ContextGraph 集成，您需要一个有效的 API 密钥。请访问 ContextGraph 官方网站注册并获取您的密钥。

To start using ContextGraph integrations, you need a valid API key. Please visit the ContextGraph official website to register and obtain your key.

总结

contextgraph-integrations 项目通过提供轻量级、非侵入式的官方集成包，显著降低了为 LangChainA framework for developing applications powered by language models through composable components. 和 CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。应用添加企业级可观测性与合规性功能的门槛。无论是为了调试复杂的代理交互，还是为了满足严格的行业监管要求，这些工具都能提供不可或缺的透明度和追溯能力。随着 AI 代理在关键业务场景中的深入应用，此类可观测性基础设施的重要性将日益凸显。

The contextgraph-integrations project significantly lowers the barrier to adding enterprise-grade observability and compliance features to LangChainA framework for developing applications powered by language models through composable components. and CrewAI一个多智能体协作框架，允许创建和管理多个AI智能体协同完成任务。 applications by providing lightweight, non-intrusive official integration packages. Whether for debugging complex agent interactions or meeting stringent industry regulatory requirements, these tools offer indispensable transparency and traceability. As AI agents become more deeply embedded in critical business scenarios, the importance of such observability infrastructure will continue to grow.

引言