AI Agent 的组成部分

有三个部分: LLM、Memory、Tools

• LLM包含推理能力,其中包含了Planning和Reflection两个能力,分解计划,以及基于结果的反思.
• Memory分为short-term and Long-term, 两个技能都具有上下文机制,区分其方法就是是否可以跨会话
• Tools,也就是可以调用外部工具,因为LLM只有推理和思考能力,但是加上tools的话就可以调用多个工具帮助其进行思考和总结甚至做出决策.

Agent WorkFlow

Conceptions

WorkFlow 分为两类: 非代理型工作流以及代理型工作流 那么其中非代理型工作流仅仅依赖LLM能力,不具备自主决策以及任务执行能力。 而代理型工作流则具备自主决策,可以自主利用工具进行搜索等能力来进行决策,以及LLM在其起到了推理作用,让传统工作流更具备响应性以及自主进化能力

Component

从图中我们可以看到三个workflow

• 那么其中automated worklfow 仅仅是一个按步骤执行的一个flow,这就好比我们的脚本一样,提前设定好,然后跑起来
• 而非代理型的workflow我们则可以看见只是将automated workflow的特定step交给LLM去接管.
• Agentic workflow可以自我抉择,监督以及得到反馈后自我寻找最优解然后执行

Agentic Workflow vs Agentic Architecture

• 代理工作流 关注任务执行的流程，即代理如何分解任务、使用工具、调整策略等。
• 代理架构 关注系统设计，即如何构建一个包含代理、工具和记忆系统的整体框架。

workflow model

planning Pattern

The panning pattern can divide complex tasks into pieces by pieces to let LLM more productively handle the task and also it can decrease the phenomenon of Hallucination.

Tools Use pattern

As we mentioned above, LLM is a reasoning model, which can not search information in realtime and check the fact whether it is true or not . So, in response to this problem, researchers introduced Retrieval-Augmented Generation (RAG), which can promote the accuracy of LLM in response.

regarding to this patten, it is perfectly applied for tasks which need realtime data support and other resources in internet

Reflection Pattern

Refection Pattern is a self-feedback mechanism, which an agent iteratively evaluates the quality of its own outputs or decisions before taking a final action.

Agentic RAG

The whole name of RAG is Retrieval-Augmented Generation Somewhere this concept had been recognized as a useless tools, But today we still need to know how it work specific content we will open a new page to discuss it.

Agentic Research Assistants

ARA(Agentic Research Assistants) is called “Deep Research” by Ai company, which is based on RAG This assistants can generate in-depth reports and detailed insights into complex topics.