什么是 Agentic Workflows?

date
Mar 23, 2025
slug
what-are-agentic-workflows-en
status
Published
tags
AI
Agent
AgenticWorkflows
summary
type
Post

中文版本

标题:什么是 Agentic Workflows?模式、用例、示例及更多 | Weaviate
发布时间:2025-03-06T00:00:00.000Z
 
notion image
notion image
AI Agents,Agentic AI,Agentic Architectures,Agentic Workflows。Agents 无处不在,但它们到底是什么?它们真的能做什么吗?
新技术带来了混杂的术语、不着边际的幻想和自封的各类专家。在这篇文章中,我们揭开了围绕 AI Agents 的噪音和炒作,解释并说明 Agentic AI 的一个关键原则:Agentic Workflows。
Agents 如果仅靠自身并不能有很大的作用。它们需要被赋予角色(Role)、目标(Goals)和结构(Structure)来实现它们的目标。这就是 workflows 的用武之地。理解 Agentic Workflows 可以帮助你理解 AI Agents 是如何运作的以及为什么是这么运作的。为了帮助你达到这一目标,我们将介绍AI Agents的关键组成部分,给出 Agentic Workflows 的简明定义,解释什么使一个工作流程成为 Agentic,详细阐述 Agentic Workflows 中的关键重复模式,详述现实世界的示例和用例,并诚实地概述使用 Agentic Workflows 的好处和挑战。
什么是AI Agents?AI Agents 是结合了大模型进行推理和决策以及现实世界交互工具的系统,使它们能够在有限的人类参与下完成复杂任务。AI Agents 被赋予特定角色,并拥有不同程度的自主权以实现其最终目标。它们还配备了记忆功能,使它们能够从过去的经验中学习,并随着时间的推移提高性能。
如需更深入了解AI Agents 、其历史以及构建它们的工具,请查看我们最近的博客文章《AI Agents 简化:在人工智能背景下的含义》。
为了更好地理解AI Agents 如何融入 Agentic Workflows,我们将探讨AI Agents 的核心组件。

AI Agents 的组件

尽管AI Agents被设计用于半自主决策,但它们依赖于一个更大的组件框架来正常运行。这个框架包括使AI Agents 能够有效推理的大模型、帮助AI Agents 完成任务的工具,以及让AI Agents 从过去的经验中学习并随着时间的推移改进响应的记忆。
notion image

推理 Reasoing

AI Agents 之所以如此有效,部分原因在于它们具备迭代推理的能力,这本质上使AI Agents 能够在整个问题解决过程中主动 “思考”。AI Agents 的推理能力源自其底层的大模型,并具有两个主要功能:规划(Planning)反思(Reflecting)
规划阶段,AI Agents 执行任务分解,即将更复杂的问题拆分为更小、可操作的步骤。这种技术使AI Agents 能够系统性地处理任务,并允许它们为不同任务使用不同的工具。它还允许进行查询分解,即将复杂查询拆分为更简单的查询,从而提高大模型响应的准确性和可靠性。
AI Agents 还通过反思其行动的结果来进行推理。这使它们能够根据从外部来源获取的结果和数据,评估并迭代调整其行动计划。大模型(LLMs)拥有静态的参数化知识,这意味着它们的理解仅限于训练期间编码的信息。为了扩展其能力,使其超越原始数据集,AI Agents 可以利用外部工具,如网络搜索引擎、API、数据库和计算框架。这意味着AI Agents 可以访问实时的外部数据,以指导其决策并完成需要与其他应用程序交互的任务。
工具通常与权限配对,例如查询 API、发送消息或访问特定文档或数据库模式的能力。下表列出了几种常见的 AI Agents工具及其执行的任务。
工具
任务
互联网搜索
检索并总结实时信息。
向量搜索
检索并总结外部数据。
代码解释器
迭代运行AI Agents 生成的代码。
API
检索实时信息并与外部服务和应用程序执行任务。
当大模型选择一个工具来帮助完成任务时,它会进行一种称为函数调用(Function Call)的行为,这扩展了其能力,使其不再局限于简单的文本生成,而是能够与现实世界进行交互。
使用哪个工具的选择可以由最终用户预先确定,也可以交给 AI Agents 来决定。让AI Agents 动态选择工具可能有助于解决更复杂的任务,但对于较简单的工作流程,使用预定义的工具可能更为高效,而动态选择工具则可能增加不必要的复杂性。

记忆 Memory

从过去的经验中学习并记住行动发生的背景,是使 Agentic Workflow 与纯粹由大模型驱动的工作流区别开来的部分原因。记忆是一个关键组件,能够跨多个用户交互和会话捕获和存储上下文和反馈。AI Agents 有两种主要类型的记忆:短期记忆(short-term)和长期记忆(long-term)。
短期记忆存储更即时的信息,如对话历史,这有助于 AI Agents 确定下一步该采取什么行动来完成其总体目标。长期记忆存储随着时间的推移在多个会话中积累的信息和知识,使得 AI Agents 能够个性化,并随着时间的推移提高性能。

什么是 Agentic Workflows?

一般来说,工作流(workflow) 是一系列相互连接的步骤,旨在完成特定的任务或目标。最简单的工作流是确定性的,意味着它们遵循预定义的步骤序列,无法适应新的信息或变化的条件。例如,一个自动化的费用审批工作流可能是这样的:“如果费用被标记为‘餐饮’且金额低于 30 美元,则自动批准。”
然而,有些工作流程会利用大模型或其他机器学习模型或技术。这些通常被称为AI Workflows,可以是 Agentic 或非 Agentic 的。在非 Agentic 工作流程中,大模型会根据指令生成输出。例如,文本摘要工作流程会将一段较长的文本作为输入,提示大模型进行摘要,并简单地返回摘要结果。然而,仅仅因为一个工作流程使用了大模型,并不一定意味着它是 Agentic 的。
Agentic Workflow  是由单个 AI Agent 或一群 AI Agents 动态执行的一系列连接步骤,以实现特定任务或目标。AI Agents 被用户授予权限,使其在有限的自主权下收集数据、执行任务并做出决策,这些决策将在现实世界中执行。Agentic Workflow 还利用了AI Agents 的核心组件,包括其推理能力、通过工具与环境交互的能力以及持久记忆,从而将传统工作流程彻底转变为响应式、自适应和自我进化的过程。
notion image
 

什么让 Workflow 智能化?

当一项 AI Workflows 由一个或多个AI Agents 引导和塑造任务进展时,它便成为 Agentic Workflow。在现有的非 Agentic 工作流程中加入AI Agents ,可以创建一种混合方法,将结构化工作流程的可靠性和可预测性与大模型的智能性和适应性相结合。
Agentic Workflows 的特点在于它们能够:
  • 制定计划。 一个 Agentic Workflow 从计划节点开始。大模型通过任务分解将复杂任务拆分为较小的子任务,然后确定最佳的执行路径。
  • 使用工具执行操作。 Agentic Workflows 使用一组预定义的工具,并结合权限来完成任务并执行生成的计划。
  • 反思与迭代。 AI Agents 可以在每个步骤评估结果,根据需要调整计划,并循环往复,直到结果令人满意。
 
如你所见,我们需要区分三种类型的工作流程:传统的非 AI Workflows、非 Agentic AI Workflows和 Agentic Workflows。传统的基于规则的工作流程与 AI Workflows的区别在于,前者使用预定义的步骤,而后者使用 AI 模型来完成任务。其次,非 Agentic 和 Agentic AI Workflows的区别在于,前者使用静态的 AI 模型,而后者使用动态的AI Agents。这使得 Agentic Workflow 比非 Agentic Workflow 更具适应性和动态性。

Agentic Architectures 与 Agentic Workflow 的区别

随着任何新兴技术的出现,都会涌现出大量新术语。虽然有些人可能会将 “Agentic 架构( Agentic Architectures)” 和 “Agentic Workflows” 这两个术语互换使用,但它们实际上有一个重要的区别。Agentic Workflow  是指AI Agents 为实现特定目标而采取的一系列步骤。这些步骤可能包括使用大模型制定计划、将任务分解为子任务、使用互联网搜索等工具完成任务,以及使用大模型反思任务结果并调整整体计划。
另一方面,Agentic Architecture 是用于实现给定任务的技术框架和整体系统设计。Agentic Architecture 多种多样且富有创意,但始终包含至少一个具有决策和推理能力的AI Agents 、用于实现目标的工具,以及短期和长期记忆系统。
 

Agentic Workflow 的集中模式

回想一下,Agentic Workflow 是完成特定任务(也称为最终目标)所采取的结构化步骤序列。因此,当我们谈论 Agentic Workflow 时,我们谈论的是使 AI Agents 能够实现其最终目标的特定行为模式。正如我们之前提到的,AI Agents 的核心组件在 Agentic Workflow 模式中起着关键作用。AI Agents 的推理能力促进了规划和反思模式,而它们使用工具与环境交互的能力则构成了工具使用模式的基础。

规划模式 Planning Pattern

规划设计模式允许 AI Agents 自主地将更复杂的任务分解为一系列更小、更简单的任务,这一过程被称为任务分解。任务分解能够带来更好的结果,因为它减少了大模型的认知负担,提高了推理能力,并最小化了幻觉和其他不准确性。
当实现最终目标的方法不明确,并且在问题解决过程中适应性至关重要时,规划尤其有效。例如,一个被指示修复软件漏洞的 AI Agents 可能会使用规划模式将任务分解为子任务,如阅读漏洞报告、识别相关代码部分、生成潜在原因列表,最后选择特定的调试策略。如果第一次修复漏洞的尝试不成功,AI Agents 可以在执行后读取错误信息并调整其策略。虽然规划可以帮助 AI Agents 更好地处理更复杂的任务,但与更确定性的工作流程相比,它也可能导致结果不那么可预测。因此,最好只在需要深入问题解决和多步推理的任务中使用规划模式。
notion image

工具使用模式 Tool Use Pattern

生成式大模型的一个显著限制是它们依赖于预训练数据,这意味着它们无法检索实时信息或验证超出其先前学习范围的事实。因此,当它们不知道答案时,可能会生成不真实的回答或 “猜测”,即产生幻觉。检索增强生成(RAG)通过为大模型提供相关的实时外部数据,帮助缓解这一限制,从而生成更准确且基于上下文的回答。
然而,工具使用超越了简单的 RAG,它允许大模型与现实世界进行动态交互,而不仅仅是从中检索数据。在 Agentic Workflow 中,工具使用模式通过允许AI Agents 与外部资源和应用程序、实时数据或其他计算资源进行交互,扩展了AI Agents 的能力。
常用工具包括 API、信息检索(例如向量搜索)、网络浏览器、机器学习模型和代码解释器。这些工具用于执行特定任务,如搜索网页、从外部数据库检索数据或读取或发送电子邮件,以帮助AI Agents 实现其目标。
notion image
notion image

反思模式 Reflection Pattern

反思是一种强大的AI Agents 设计模式,相对容易实现,并且可以显著提升 Agentic Workflows 的改进效果。反思模式是一种自我反馈机制,AI Agents 在最终确定响应或采取进一步行动之前,会迭代评估其输出或决策的质量。这些评估结果随后被用来优化AI Agents 的方法,纠正错误,并改进未来的响应或决策。
当AI Agents 不太可能在第一次尝试中就成功实现其目标时(例如编写代码),反思尤其有用。在这种情况下,AI Agents 可能会生成一段代码,在沙盒或执行环境中运行它,并将错误信息迭代反馈给大模型,同时提供改进代码的指令,直到代码成功执行为止。
反思的力量在于 AI Agents 能够评估自己的输出,并将这些洞察动态地整合到工作流程中,从而在没有直接人类反馈的情况下实现持续改进。这些反思可以被编码到AI Agents 的记忆中,从而在当前用户会话中能够更高效地解决问题,并通过适应用户偏好来实现个性化,从而改善未来的交互。
notion image

Agentic Workflows 的应用场景

原子设计模式,如规划和工具使用,可以通过创造性的方式组合,以有效利用 Agentic AI 在多个领域中完成各种任务。除了组合设计模式外,还可以为 AI Agents 提供不同的工具组合,甚至赋予它们根据需要动态选择工具的能力。它们还可以与人类反馈循环集成,并赋予不同水平的自主权和决策权。
这些多样化的配置使得AI Agents 工作流能够针对各行业的广泛任务进行定制。为了展示这一点,我们概述了两个特别强大的用例:AI Agents RAG 和 AI Agents 研究助手。

Agentic RAG

检索增强生成(RAG) 是一个通过从外部数据源检索相关数据来增强大模型知识的框架。Agentic RAG 则在 RAG 流程中引入了一个或多个 AI Agents 。
在规划阶段,AI Agents 可以通过查询分解将复杂查询拆分为较小的子查询,或者确定是否需要向用户请求更多信息以完成请求。
AI Agents还可以用于评估检索到的数据和响应的相关性和准确性,然后再将其传递给用户。如果响应没达到用户预期,AI Agents 可以重新制定查询,重新进行查询分解步骤,甚至创建一个新的响应计划。像这样的 Agentic Workflow 可以通过不同的 Agentic 架构来构建。
notion image

研究助手 Agentic Research Assistants

AI Agents 研究助手,也被一些 AI 公司称为 “Deep Research”,通过浏览网络和各种外部数据生成关于复杂主题的深度报告和详细见解。这些助手利用 Agentic RAG 从网络和其他外部来源检索信息以响应用户查询。然而,这些助手与传统 RAG 的不同之处在于它们能够整合和分析信息,而不仅仅是从外部来源检索相关数据来增强大模型生成的输出。
这种独特的能力归功于几个特点。首先,AI Agents 研究助手通常使用专门针对网页浏览、任务分解和动态规划进行微调的大模型。其次,这些工作流程中的 AI Agents 会主动寻求用户指导,请求更多信息或澄清以更好地理解最终目标。第三,这些AI Agents 能够根据检索到的信息调整计划并改变方向。这意味着在合成新信息时,它们可以追求新的、有趣的角度,并连续查询多个数据源,直到获得必要的数据。因此,AI Agents 研究助手能够获得更深入的洞察,识别随时间变化的趋势,并编译完整的主题报告,而不仅仅是检索现有知识。在撰写本文时,OpenAIPerplexityGoogle都公开提供了各自版本的深度研究功能。

编程助手 Agentic Coding Assistants

AI Agents 编程助手可以在最少人工干预的情况下生成、重构、优化和调试代码。非AI Agents 编程助手,如第一版的 GitHub Copilot,由经过微调的生成式大模型驱动,专门用于生成代码,但其功能仅限于此 —— 生成代码。
使编程助手具备AI Agents 特性的,是其能够通过执行生成的代码与环境交互,并根据执行结果、错误或反馈进行迭代优化的能力。这些助手还可以被授予权限,通过创建提交和 PR 来对现有代码库进行更改,例如 Anthropic 的 Claude Code,这是自动化软件开发过程中的重要一步。AI Agents 编程助手还可以用于建议终端命令以及其他代码的修改和添加,并在执行前等待明确的人工批准,例如 Cursor 的 Agent,这使得人类能够完全控制 AI Agents 。此外,重要的是,AI Agents 编程助手能够通过将错误编码到长期记忆中,从错误中学习,从而随着时间的推移变得更聪明。
 

Agentic Workflow 示例

既然我们已经概述了 Agentic Workflow 的一些应用场景,接下来我们将更详细地了解两个实际AI Agents 的工作流程中的各个步骤:Claygent 和 ServiceNow AI Agents。每个工作流程都使用其独特的模式和工具组合,赋予其AI Agents 不同程度的自主性和决策能力,并依赖于不同层次的人类反馈和参与。

Claygent (Clay)

潜在客户研究和数据丰富对于增长和销售团队来说可能是一项繁琐的任务。Clay 是一家数据丰富和外展自动化公司,通过 Claygent 简化了这一流程。Claygent 是一个由人工智能驱动的研究 AI Agents ,持续扫描网络和内部数据库,以提供实时的、可操作的见解。
假设你想使用 Claygent 根据一系列姓名和电子邮件地址来丰富 LinkedIn 个人资料,然后发送个性化的介绍信息。首先,你需要指定你想要查找的数据字段(例如工作经验、教育背景、技能),这些字段会被注入到一个预配置的提示模板中。AI Agents 的大模型会处理查询,使用网络爬虫工具在网络上搜索 LinkedIn 的 URL,并从 LinkedIn 个人资料中提取所需的数据。然后,这些数据可以被发送到另一个大模型,你可以指示它以你希望的方式总结或分析这些丰富的数据。接着,可以使用同一个大模型(或另一个)为每个个人资料创建个性化的联系信息。
Claygent 是一个相对灵活的 Agentic Workflow 示例,它可以以创造性的方式进行定制,同时仍然通过为特定任务预配置的提示模板为AI Agents 提供指导。

ServiceNow AI Agents

ServiceNow 是一个基于云的平台,旨在简化和自动化跨 IT、运营、人力资源和客户服务领域的工作流程。他们的 ServiceNow 平台现在集成了AI Agents ,旨在自动化重复性任务和现有工作流程,同时仍由人类完全掌控决策权。
这是一个关于 Agentic Workflow 如何帮助解决技术支持案例的示例。当客户提交技术支持工单时,该工作流程被触发。工单中的信息随后传递给一个或多个AI Agents ,这些 AI Agents 在内部 IT 支持知识库上执行 RAG(检索增强生成)。AI Agents 总结发现的内容,分析类似案例,并为 IT 支持专家生成摘要。最后,它生成关于如何继续进行的建议,专家可以批准或拒绝该建议。
ServiceNow AI Agents 代表了一种在生产环境中部署 AI Agents 的创新但更为谨慎的方法,赋予它们严格的角色和任务,并在影响最终用户或客户的决策上给予有限(如果有的话)的自主权。

Agentic Workflows 的益处与局限性。

AI Agents 已经迅速从机器学习社区走向主流。鉴于围绕 Agentic AI 的所有兴奋、期待和期望,可能很难区分炒作与现实,并理解其真正的能力和局限性。在本节中,我们将为您提供一个关于 Agentic Workflow 的优势、挑战和局限性的较为中立的观点。

Agentic Workflows 的优势

Agentic Workflows 通过使AI Agents 能够规划、适应并随时间迭代,超越了传统的自动化。与遵循固定规则的确定性工作流不同,Agentic Workflows 能够动态应对复杂性,通过反馈优化其方法,并扩展以处理更高级的任务。这种适应性使它们在需要灵活性、学习和决策的场景中尤为有价值。
让我们更详细地看看 Agentic Workflows 的优势:
  • 灵活性、适应性和可定制性。 静态的确定性工作流难以适应不断变化的情况和意外困难。而 Agentic Workflows 则提供了根据任务难度进行调整和演变的灵活性,确保它们始终保持相关性并提供最佳解决方案。它们还可以通过组合不同的模式进行定制,实现模块化设计,从而在需求和复杂性增长时进行迭代升级。
  • 在复杂任务上的性能提升。 通过将复杂任务分解为更小、可管理的步骤(通过任务分解和规划),Agentic Workflow 显著优于确定性的零样本方法。
  • 自我纠正和持续学习。 反思模式使 Agentic Workflow 能够评估自身行为,完善自身策略,并随着时间的推移迭代自身。通过利用短期和长期记忆,它们从过去的经验中学习,在每次迭代中变得更有效和个性化。
  • 操作效率和可扩展性。 Agentic Workflow 可以高精度地自动化重复任务(如果构建得当),在特定场景中减少人工操作和运营成本。它们还可以轻松扩展,使其成为处理更大工作量或复杂系统的理想选择。
请记住,AI Agents 仍然是一项新兴技术,随着研究人员和用户发现将AI Agents 融入工作流程的新方法,这份优势列表很可能会继续扩展。

Agentic Workflows 的挑战与局限

尽管AI Agents具有诸多优势和创新特性,但它们也带来了一些挑战和局限。由于其概率性本质,AI Agents 天生会增加工作流程的复杂性。而且,尽管 AI Agents 可以用于自动化流程,并不意味着在它们在任何情况下都应该被使用。以下是 Agentic Workflows 中最显著的挑战和局限:
  • 简单任务的过度复杂化。 在处理诸如表单填写或基本数据提取等直接工作流时,AI Agents 可能会增加额外开销。在确定性、基于规则的自动化已经足够的情况下,引入AI Agents 可能会导致效率低下、成本增加,甚至可能降低性能。
  • 自主性增加导致可靠性降低。 随着 AI Agents 在工作流中获得更多决策权,其概率性本质可能引入不可预测性,使得输出结果可靠性降低且更难控制。为 AI Agents 实施并积极维护防护措施,并持续审查其被授予的权限至关重要。
  • 伦理和实际考量。 并非所有决策都应委托给 AI 系统。在高风险或敏感领域使用 AI Agents 需要谨慎监督,以确保负责任地部署并防止意外后果。
鉴于这些限制,我们建议花时间反思在特定工作流程中使用 AI Agents 是否真正必要。以下问题可能有助于您做出判断:
  • 任务是否足够复杂,需要适应性决策,还是确定性方法就足够了?
  • 更简单的 AI 辅助工具(如不带AI Agents 的 RAG)能否达到相同效果?
  • 工作流程是否涉及不确定性、变化的条件或多步推理,AI Agents 能否更有效地处理?
  • 赋予AI Agents 自主性会带来哪些风险,是否可以缓解?

总结

Agentic Workflows 是强大的工具,能够帮助自动化完成需要决策和推理的复杂任务。在本文中,我们回顾了AI Agents 的核心组成部分,包括记忆、工具和推理能力,以及它们如何为 Agentic Workflows 做出贡献。我们还讨论了常见的工作流模式,如规划、工具使用和反思,这些模式可以单独或组合使用,以创建动态的工作流。此外,我们概述了两个特别有效的用例,Agentic RAG 和 Agentic research agents,并描述了市场上已有的两个AI Agents ——Clay 的 Claygent 和 ServiceNow 的 AI Agents 的工作流程。最后,我们探讨了 Agentic Workflows 的优势,以及它们的局限性和挑战。AI Agents 背后的技术以及我们对它们的理解都在不断发展。本文旨在让你对AI Agents 在工作流程中的功能有一个基本的了解,但绝不是对这一主题的详尽探讨。
 
 

英文原文

Title: What Are Agentic Workflows? Patterns, Use Cases, Examples, and More | Weaviate
Published Time: 2025-03-06T00:00:00.000Z
Markdown Content:
notion image
AI agents. Agentic AI. Agentic Architectures. Agentic Workflows. Agents are everywhere. But what are they really? And can they actually do anything?
New technology brings with it a muddled mixture of confusing terminology, wild expectations, and self-proclaimed online experts. In this article, we cut through the noise and hype surrounding AI agents to explain and illustrate a critical tenet of Agentic AI: Agentic Workflows.
Agents, completely on their own, can’t do much. They need to be given roles, goals, and structure to achieve their goals. This is where workflows come in.
Understanding Agentic Workflows allows you to understand how and why AI agents operate as they do. To help you get there, we’ll go through the key components of AI agents, give you a concise definition of Agentic Workflows, explain what makes a workflow Agentic, elaborate on key recurring patterns in Agentic Workflows, detail real-world examples and use cases, and give an honest overview of the benefits and challenges to using Agentic Workflows.

What are AI agents?

AI agents are systems that combine LLMs for reasoning and decision-making with tools for real-world interaction, enabling them to complete complex tasks with limited human involvement. Agents are assigned specific roles and given varying degrees of autonomy to accomplish their end goal. They are also equipped with memory, allowing them to learn from past experiences and enhance their performance over time.
For a more in-depth explanation of AI agents, their history, and tools for building them, check out our recent blog post, Agents Simplified: What we mean in the context of AI.
To better understand how AI agents fit into Agentic Workflows, we’ll explore the core components of AI agents.

Components of AI Agents

Although AI agents are designed for semi-autonomous decision-making, they rely on a larger framework of components to function properly. This framework consists of LLMs that enable the agent to reason effectively, tools that help the agent complete its tasks, and memory that allows the agent to learn from past experiences and improve responses over time.
notion image

Reasoning

Part of what make AI agents so effective is their capacity for iterative reasoning, essentially allowing the agent to actively “think” throughout the entire problem-solving process. The reasoning capabilities of an AI agent stem from its underlying LLM and serve two primary functions: planning and reflecting.
In the planning phase, the agent performs task decomposition, the process of breaking down a more complex problem into smaller, actionable steps. This technique allows for agents to approach tasks systematically and allows them to use different tools for different tasks. It also allows for query decomposition, in which complex queries are broken down into simpler queries, which improves the accuracy and reliability of responses from the LLM.
Agents also reason through reflecting on the outcomes of their actions. This allows them to evaluate and iteratively adjust their plan of action based on results and data pulled from external sources.

Tools

LLMs possess static, parametric knowledge, meaning their understanding is confined to the information encoded during training. To expand their capabilities beyond their original dataset, agents can leverage external tools, like web search engines, APIs, databases, and computational frameworks. This means that the agent has access to real-time external data to guide its decision-making and accomplish tasks that require it interact with other applications.
Tools are often paired with permissions, such as the ability to query APIs, send messages, or access specific documents or database schemas. The table below outlines several common tools for AI agents along with the tasks they perform.
Tool
Task
Internet search
Retrieve and summarize real-time information.
Vector search
Retrieve and summarize external data.
Code interpreter
Iteratively run code generated by agents.
API
Retrieve real-time information and perform tasks with external services and applications.
When the LLM selects a tool to help achieve a task, it engages in a behavior called function calling, extending its capabilities beyond simple text generation and allowing it to interact with the real-world.
The choice of which tool to use can be predetermined by the end user or be left to the agent. Letting the agent dynamically select tools can be helpful for solving more complex tasks but can add unnecessary complexity for simpler workflows, when predefined tools would be more efficient.

Memory

Learning from past experiences and remembering the context in which actions take place are part of what set Agentic Workflows apart from purely LLM-driven workflows. Memory is a key component that enables the capture and storage of context and feedback across multiple user interactions and sessions. Agents have two main types of memory: short-term memory and long-term memory.
Short-term memory stores more immediate information like conversation history, which helps the agent determine which steps to take next to complete its overall goal. Long-term memory stores information and knowledge accumulated over time, throughout multiple sessions, allowing for personalization of the agent and improved performance over time.

What are Agentic Workflows?

In general, a workflow is a series of connected steps designed to achieve a specific task or goal. The simplest types of workflows are deterministic, meaning they follow a predefined sequence of steps and are unable to adapt to new information or changing conditions. For example, an automated expense approval workflow could look like this: “if expense is tagged as ‘Food and Meals’ and is less than $30, automatically approve.”
Some workflows, however, leverage LLMs or other machine learning models or techniques. These are often referred to as AI workflows, and can be Agentic or non-Agentic. In a non-Agentic Workflow , a LLM is prompted with an instruction and generates an output. For example, a text summarization workflow would take a longer passage of text as its input, prompt a LLM to summarize it, and simply return the summary. However, just because a workflow uses a LLM, doesn’t necessarily mean that it’s Agentic.
An Agentic Workflow is a series of connected steps dynamically executed by an agent, or series of agents, to achieve a specific task or goal. Agents are granted permissions by their users, which give them a limited degree of autonomy to gather data, perform tasks, and make decisions to be executed in the real-world. Agentic Workflows also leverage the core components of AI agents including, their capacity for reasoning, ability to use tools to interact with their environment, and persistent memory to completely transform traditional workflows into responsive, adaptive, and self-evolving processes.
notion image

What makes a workflow Agentic?

An AI workflow becomes Agentic when one or more agents guide and shape the progression of tasks. Adding agents to an existing non-Agentic Workflow creates a hybrid approach that combines the reliability and predictability of structured workflows with the intelligence and adaptability of LLMs. Agentic Workflows are defined by their ability to:
  • Make a plan. An Agentic Workflow starts with planning. The LLM is used to break down complex tasks into smaller sub-tasks through task decomposition and then determines the best execution route.
  • Execute actions with tools. Agentic Workflows use a set of predefined tools paired with permissions in order to accomplish tasks and carry out their generated plan.
  • Reflect and iterate. Agents can assess results at each step, adjust the plan if needed, and loop back until the outcome is satisfactory.
As you can see, we need to differentiate between three types of workflows: traditional non-AI workflows, non-Agentic AI workflows, and Agentic Workflows. The difference between a traditional, rule-based workflow and an AI workflow is the use of predefined steps vs. the use of AI models to accomplish a task. Second, the difference between non-Agentic and Agentic AI workflows is the use of static AI models vs. dynamic AI agents. This makes the Agentic Workflow more adaptive an dynamic than a non-Agentic Workflow .

The difference between Agentic Architectures and workflows

With any emerging technology, comes a flood of new terminology. While some may use the terms “Agentic Architectures” and “Agentic Workflows” interchangeably, they actually have an important distinction.
An Agentic Workflow , is the series of steps taken by an agent to achieve a certain goal. These steps may include using LLMs to create a plan, break down tasks into subtasks, using tools like internet search to accomplish tasks, and using LLMs to reflect on the outcomes of tasks and adjust their overall plan.
An Agentic Architecture, on the other hand, is the technical framework and overall system design used to achieve a given task. Agentic Architectures are diverse and creative but always contain at least one agent with decision-making and reasoning capabilities, tools the agent can use to accomplish its goals, and systems for short-term and long-term memory.
note
Explore the most powerful Agentic Architectures, visually illustrated for instant understanding. Download the free e-book here.

Patterns in Agentic Workflows

Recall that an Agentic Workflow is the structured series of steps taken to complete a specific task, also known as a final target. So when we talk about Agentic Workflows, we talk about specific patterns of behavior that enable agents to achieve their final target. The core components of AI agents, as we mentioned earlier, play a key role in Agentic Workflow patterns. The capacity for agents to reason facilitates both the planning and reflection patterns, while their ability to use tools to interact with their environment underlies the tool use pattern.

Planning Pattern

The planning design pattern allows agents to autonomously break down more complex tasks into series of smaller and simpler tasks, a process known as task decomposition. Task decomposition leads to better results because it reduces the cognitive load on the LLM, improves reasoning, and minimizes hallucinations and other inaccuracies.
Planning is especially effective when the method to achieve a final target is unclear and adaptability in the problem solving process is paramount. For instance, an AI agent instructed to fix a software bug would likely use the planning to pattern to break down the task into subtasks like reading the bug report, identifying the relevant code sections, generating a list of potential causes, and finally selecting a specific debugging strategy. If the first attempt to fix the bug doesn’t work, the agent can read the error messages after execution and adapt its strategy.
While planning can help agents better tackle more complex tasks, it can also lead to less predictable results than more deterministic workflows. As a result, it’s best to only use the planning pattern with tasks that require intense problem-solving and multi-hop reasoning.
notion image

Tool Use Pattern

A significant constraint of generative LLMs is their reliance on pre-existing training data, meaning they cannot retrieve real-time information or verify facts beyond what they have previously learned. As a result, they may generate non-factual responses or “guess” when they don’t know the answer. Retrieval Augmented Generation (RAG) helps mitigate this limitation by providing the LLM with relevant, real-time external data, enabling more accurate and contextually grounded responses.
Tool use, however, goes beyond naive RAG by allowing the LLM to dynamically interact with the real world, as opposed to simply retrieving data from it. In Agentic Workflows, the tool use pattern expands the capabilities of agents by allowing them to interact with external resources and applications, real-time data, or other computational resources.
Common tools include APIs, information retrieval (e.g. vector search), web browsers, machine learning models, and code interpreters. These tools are used to perform specific tasks, like searching the web, retrieving data from an external database, or reading or sending emails that help the agent achieve their target.
notion image

Reflection Pattern

Reflection is a powerful Agentic design pattern that is relatively simple to implement and can lead to significant gains in improvement for Agentic Workflows. The reflection pattern is a self-feedback mechanism in which an agent iteratively evaluates the quality of its outputs or decisions before finalizing a response or taking further action. These critiques are then used to refine the agent's approach, correct errors, and improve future responses or decisions.
Reflection is particularly useful when the agent is unlikely to succeed in accomplishing its target goal on the first attempt, such as writing code. In this case, an agent may generate a code snippet, run it in a sandbox or execution environment, and iteratively feed errors back into the LLM with instructions to refine the code until it executes successfully.
The power of reflection lies in the agent’s ability to critique its own outputs and dynamically integrate those insights into the workflow, enabling continuous improvement without direct human feedback. These reflections can be encoded in the agent’s memory, allowing for more efficient problem-solving during the current user session and enabling personalization by adapting to user preferences and improve future interactions.
notion image

Agentic Workflows Use Cases

Atomic design patterns, like planning and tool use, can be combined in creative ways to effectively leverage Agentic AI for a variety of tasks across diverse domains. In addition to combining design patterns, AI agents can be provided with different combinations of tools and even be granted the ability to dynamically select tools as needed. They can also be integrated with human feedback loops and given varying levels of autonomy and decision-making powers.
These diverse configurations allow Agentic Workflows to be tailored for a wide range of tasks across industries. To demonstrate this, we outline two especially powerful use cases: Agentic RAG and Agentic research assistants.

Agentic RAG

Retrieval-Augmented Generation (RAG) is a framework that augments the knowledge of a LLM by providing it with relevant data retrieved from an external data source. Agentic RAG incorporates one or more agents into the RAG pipeline.
During the planning phase, an agent can break down complex queries into smaller subqueries through query decomposition or determine whether it needs to ask the user for additional information to complete the request.
An AI agent can also be used to evaluate the relevance and accuracy of retrieved data and responses before it’s passed on to the user. If the response is not satisfactory, the agent can reformulate the query, revisit the query decomposition step, or even create a new plan for responding to the query.
notion image
note
Agentic Workflows like this one can be built with different Agentic Architectures. If you are curious about potential architectures for the above workflow, download our free e-book on Agentic Architectures!

Agentic Research Assistants

Agentic research assistants, also referred to as “deep research” by some AI companies, generate in-depth reports and detailed insights on complex topics by scouring the web and all sorts of external data. These leverage Agentic RAG to retrieve information from the web and other external sources in response to user queries. What sets these assistants apart from traditional RAG, however, is their ability to synthesize and analyze information, as opposed to simply retrieving relevant data from external sources to enhance the output generated by a LLM.
This unique ability is attributed to a few features. First, Agentic research assistant generally use LLMs that have been fine-tuned specifically for web browsing, task decomposition, and dynamic planning. Second, agents in these workflows actively seek user guidance, requesting additional information or clarification to better understand the final goal. Third, these agents are able to adapt their plans and change course depending on the information they retrieve. This means that they can pursue new, interesting angles when synthesizing novel information and query multiple data sources consecutively until they get the necessary data.
As a result, Agentic research assistants are able to gain deeper insights, identify trends over time, and compile full reports on topics as opposed to simply retrieving existing knowledge. At the time of writing, OpenAI, Perplexity, and Google all have their own version of Deep Research publicly available.

Agentic Coding Assistants

Agentic coding assistants can generate, refactor, refine, and debug code with minimal human intervention. Non-Agentic coding assistants, like the first version of GitHub Copilot, are powered by generative LLMs fine-tuned to generate code, but are limited to doing just that — generating code.
What makes a coding assistant Agentic is its ability to interact with its environment by executing generated code and iteratively refine it based on execution results, errors, or feedback. These assistants can also be enabled with permissions to make changes to an existing code base by creating commits and PRs, like Anthropic’s Claude Code, an important step in automating the software development process. Agentic coding assistants can also be used to suggest terminal commands and other code changes and additions and wait for explicit human approval before execution, like Cursor’s Agent, allowing humans to be fully in control of the agent. Additionally, and importantly, Agentic coding assistants can learn from their mistakes by encoding them in long-term memory, allowing them to grow smarter over time.

Agentic Workflow Examples

Now that we’ve outlined some use cases for Agentic Workflows, we’ll take a more detailed look at the individual steps of the workflows of two real-world agents: Claygent and ServiceNow AI Agents. Each workflow uses its own unique combination of patterns and tools, gives its agents varying degrees of autonomy and decision-making capabilities, and relies on different levels of human feedback and involvement.

Claygent (Clay)

Lead research and data enrichment can be a tedious task for growth and sales teams. Clay, a data enrichment and outreach automation company, streamlines this process with Claygent—an AI-powered research agent that continuously scans the web and internal databases to deliver real-time, actionable insights.
Say you want to use Claygent to enrich a LinkedIn profile based on a list of names and email addresses and then send a personalized introduction message. First, you specify the fields of data you’re looking for (e.g. work experience, education, skills), which is injected into a pre-configured prompt template. The agent’s LLM processes the query, uses a web scraping tool to scour the web for a LinkedIn URL, and extracts the desired data from the LinkedIn profile. This data can then be sent to another LLM that you can instruct to summarize or analyze the enriched data however you want. The same LLM (or a different one) can then be used to create a personalized outreach message for each profile.
Claygent is an example of a relatively flexible Agentic Workflow that can be customized in creative ways, while still providing guidance to agents through pre-configured prompt templates for specific tasks.

ServiceNow AI Agents

ServiceNow is a cloud-based platform that streamlines and automates workflows across IT, operations, HR, and customer service domains. Their ServiceNow Platform now includes access to AI agents, intended to automate repetitive tasks and pre-existing workflows, while still leaving humans in full control of making decisions.
Here’s an example of how an Agentic Workflow can help resolve a technical support case. The workflow is triggered when a customer submits a ticket for technical support. The information from the ticket is then passed to one or more agents that perform RAG on an internal IT support knowledge base. The agent summarizes the findings, analyzes similar cases, and generates a summary for the IT support specialist. Finally, it generates a recommendation for how to proceed, which the specialist can either approve or deny.
ServiceNow AI Agents represent an innovative but more cautious approach to deploying agents in production, giving them strict roles and tasks to accomplish and limited, if any, autonomy to make decisions that affect the end user or customer.

Benefits and Limitations of Agentic Workflows

AI agents have rapidly moved beyond the machine learning community and into the mainstream. Given all the excitement, anticipation, and expectations around Agentic AI, it can be difficult to separate hype from reality and understand its true capabilities and limitations. In this section, we give you a balanced view of the benefits, challenges, and limitations of Agentic Workflows.

Benefits of Agentic Workflows

Agentic Workflows go beyond traditional automation by enabling AI agents to plan, adapt, and improve over time. Unlike deterministic workflows, which follow fixed rules, Agentic Workflows can dynamically respond to complexity, refine their approach through feedback, and scale to handle more advanced tasks. This adaptability makes them particularly valuable in scenarios where flexibility, learning, and decision-making are essential.
Let’s take a closer look at the benefits of Agentic Workflows:
  • Flexibility, adaptability, and customizability. Static, deterministic workflows struggle to adapt to evolving situations and unexpected difficulties. Agentic Workflows, on the other hand, offer the flexibility to adjust and evolve based on the task difficulty, ensuring they always stay relevant and give the best solution. They can also be customized by combining different patterns, enabling a modular design that allows iterative upgrades as needs and complexity grows.
  • Improved performance on complex tasks. By breaking down complex tasks into smaller manageable steps (through task decomposition and planning), Agentic Workflows significantly outperform deterministic, zero-shot approaches.
  • Self-correcting and continuous learning. The reflection pattern allows Agentic Workflow to evaluate their own actions, refine strategies, and improve outcomes over time. Utilizing both short- and long-term memory, they learn from past experiences to become more effective and personalized with each iteration.
  • Operational efficiency and scalability. Agentic Workflows can automate repetitive tasks with high accuracy (if built right), reducing manual effort and operational costs in specific scenarios. They can also scale easily, making them ideal for handling larger workloads or complex systems.
Keep in mind that AI agents are still an emerging technology, and that this list of benefits is likely to expand as researchers and users discover novel ways of incorporating agents into workflows.

Challenges and Limitations of Agentic Workflows

Despite their benefits and innovative features, AI agents also come with a number of challenges and limitations. Because of their probabilistic nature, AI agents inherently add complexity to workflows. And just because agents can be used to automate processes, doesn’t mean that they should be used. Here are a few of the most notable challenges and limitations of Agentic Workflows:
  • Unnecessary complexity for simple tasks. AI agents can add overhead when used for straightforward workflows like form entry or basic data extraction. In cases where deterministic, rules-based automation is sufficient, introducing agents may lead to inefficiencies, extra expense, and possibly reduced performance.
  • Reduced reliability as a result of increased autonomy. As agents gain more decision-making power within a workflow, their probabilistic nature can introduce unpredictability, making outputs less reliable and harder to control. Implementing and actively maintaining guardrails for agents and continually reviewing their granted permissions is critical.
  • Ethical and practical considerations. Not all decisions should be delegated to AI systems. Using agents in high-stakes or sensitive areas requires careful oversight to ensure responsible deployment and prevent unintended consequences.
Given these limitations, we recommend taking time to reflect on whether using an agent is truly necessary in a given workflow. Some questions to help you determine this may include:
  • Is the task complex enough to require adaptive decision-making, or would a deterministic approach suffice?
  • Would a simpler AI-assisted tool (such as RAG without an agent) achieve the same outcome?
  • Does the workflow involve uncertainty, changing conditions, or multi-step reasoning that an agent could handle more effectively?
  • What are the risks associated with giving the agent autonomy, and can they be mitigated?

Summary

Agentic Workflows are powerful tools to help automate the completion of complex tasks that require decision-making and reasoning. In this article, we reviewed the core components of AI agents, including, memory, tools, and reasoning capabilities and how they contribute to Agentic Workflows. We also discussed common workflow patterns, like planning, tool use, and reflection that can be used in isolation or combination to create dynamic workflows. Furthermore, we outlined two particularly effective use cases, Agentic RAG and Agentic research agents, and described the workflows behind two agents already on the market, Clay’s Claygent and ServiceNow’s AI Agents. Finally, we touched on the benefits of Agentic Workflows as well as their limitations and challenges.
The technology behind AI agents is continuously evolving, as is our understanding of them. This article is intended to give you a basic understanding how AI agents function in workflows but is by no means an exhaustive exploration of the topic.
For a more detailed view and explanation of specific Agentic Architectures, download our free e-book.
notion image

Resource guide

Ready to start building?

Check out the Quickstart tutorial, or build amazing apps with a free trial of Weaviate Cloud (WCD).

© 溪河 2021 - 2025