4597 - 超越基础 AI 代理团队：使用 CrewAI Flows 实现状态和路由.pdf

1、Building Robust AgentsWith CrewAI FlowsTony KipkemboiSr.DevRel Eng.at CrewAIWhat to expect from this workshopLearn how to build AI Agents with CrewAI frameworkBuild your first Multi Agent System!Lean how to apply these principles in your career and companiesAnatomy of a CrewCrews:Collaborative agent

2、s,more autonomous and dynamic.Better for exploratory,less deterministic tasks.Flows:Low level,event-driven,control layer.Ideal for structured,repeatable steps with tight control.80%effort:Craft clear tasks(task description&expected output)20%effort:Polish agent personasThe 80/20 Rule“Well-designed t

3、asks rescue imperfect agents”Designing Agent Role(Job Title)Specific&specialized(e.g.,“Clinical Trial Data Auditor”)Align with real-world job archetypesExpress domain expertiseAgent Goal(Brief Expectations)Outcome-focused&measurableEmbed quality standards&success criteriaWhat does“good”look like?Age

4、nt Backstory(Job Experience)Establish expertise and experienceDescribe working style&valuesEnsure consistency with role&goalSpecialists GeneralistsFocused agents deliver precise outputsExample:Technical Blog Writer Agentinstead of Writer AgentIf the goal is to have a technical blog creation agent.St

5、art a ConversationTrigger deep research?ORAnswerResearchStart a ConversationORAnswerResearchLLMTrigger deep research?LLMStart a ConversationORAnswerResearchLLMCrewTrigger deep research?LLMYou can build an agentsBut can it run reliable every time?More importantly:Can you trust it?GuardrailsHITL overs

6、ightReasoning AgentsPre and Post Hooks Flows as a control backboneHow to Build Agents you TrustIncluding Deterministic Controls to Agentic systemTaskUses actual code to checks the output and whether approve it or notCodedef count_of_words(self,result:TaskOutput)-Tuplebool,Any:tot