APPLICATION

多代理博客生成器

我用3个AI代理组成了一个团队。内容策略师负责编写大纲，博客作家负责撰写博客，编辑负责校对。

admin

Jan 17, 2025 • 6 min read

在黑客松期间，我决定探索多AI代理系统。与通常的LLM交互不同，机器需要在每一步都持续获得人类的反馈，而多AI代理系统中，多个LLM代理（具有不同的角色、目标和背景故事）相互协作并互相提示。

除此之外，这些代理可以与工具互动，如使用API、读取个人数据等。它与模型无关，因此你可以使用OpenAI GPT模型、Cohere、Llama、Huggingface等。

我用了一天时间完成了DeepLearning.ai关于多AI代理系统的快速课程，并开始动手实践。接下来的几天里，我用3个代理组成了一个团队。内容策略师负责编写大纲，博客作家负责撰写博客，编辑负责校对。

content_strategist = Agent(
  role='Senior Content Strategist',
  goal="Plan an outline for a blog post content to highlight new discoveries and educate your readers about {topic}",
  backstory="""You work for a fun startup that loves to tinker with new technologies.
  You love your work and have a passion for structured communication.
  Your expertise lies in organizing content effectively so it's easy to follow and understand.
  You have a knack for dissecting huge chunks of information, highlighting the important points, creating lesson plans, and sharing your learnings with other people.""",
  verbose=True,
  allow_delegation=False,
  tools=[search_tool, scrape_tool, read_notes, semantic_search_notes],
  step_callback=lambda step: self.step_callback(step, "Content Strategist")
)
blog_writer = Agent(
  role='Tech Blogger',
  goal='Craft a compelling blog post about {topic} based on given outline and notes',
  backstory="""You are a renowned tech blogger, known for your insightful and engaging articles that are easy to follow and learn from.
  You have years of experience writing guides about {topic}, learning reflections, teaching and mentoring people.
  You are excellent at synthesizing lots of information with clarity and brevity.
  You like to make your articles fun to read, like metaphors and appropriately inserting emojis where suitable.
  You transform complex concepts into compelling narratives that is easy to understand.""",
  verbose=True,
  allow_delegation=True,
  tools=[read_notes, semantic_search_notes],
  step_callback=lambda step: self.step_callback(step, "Tech Blogger")
)
editor = Agent(
  role='Senior Blog Editor',
  goal='Proofread blogs to ensure the content is fun to read, informative, accurately includes the important points',
  backstory="""You work as a senior blog editor with years of experience in writing.
  You have great attention to details, ensuring the concepts are best delivered, well-written, and search engine optimized.
  You take pride in only publishing the very best writing.""",
  verbose=True,
  allow_delegation=True,
  tools=[read_notes, semantic_search_notes],
  step_callback=lambda step: self.step_callback(step, "Senior Blog Editor")
)

提示感觉是另一种工艺。我不会称之为编程。正如课程中所述，一个很好的思维框架是像管理者一样思考。目标是什么？流程是什么？我需要雇用什么样的人来完成它？

这里是一个关于Crew AI的示例摘录，使用我从课程中写下的笔记。由我本人撰写，Crew AI多AI代理系统。

在Crew AI的核心是多代理系统，其中AI代理协同工作以实现共同目标。这些代理被训练成专门执行特定任务，如研究、写作或分析，并可以无缝地相互交互以实现复杂目标。通过利用Crew AI，开发人员可以创建动态系统，其中代理可以委派任务、共享信息并协同工作以实现最佳结果。

通过将任务结构化为明确的角色、目标和背景故事，Crew AI使开发人员能够创建复杂的workflow，模仿现实世界的情景。从内容规划到客户服务自动化，Crew AI可以简化流程、提高效率并增强整体用户体验。