由于 AI Agent 是一个实践性很强的新领域，同时目前可以借鉴的方法非常有限，因此我主要还是以探索为主并制定了以下学习策略：
我暂时不需要
理解Framework的具体代码架构
探索目标
选择一个相对成熟的 AI Agent 框架
采用低成本的方式进行尝试
结合自身工作，尝试在小场景中实践
要开心有趣并且享受学习过程
框架选择上，我采用了 CrewAI 作为主要开发框架，因为它封装度较高，可以让我专注于任务设计，而不是花大量时间在环境部署上。
在项目管理中的应用
工作中会涉及到战略管理的内容，常涉及年度规划、目标拆解和战略项目规划等，这些任务通常耗费大量时间和人力。若能优化流程、加快讨论方案的产出，不仅能显著提升团队运作效率，还能确保战略目标更精准、高效地落地。
在实际项目中，我发现这些看似复杂的任务，也是有迹可循的。对应的角色的分工也相对明确，因此我认为 AI Agent 可能在这些场景中有所帮助。因此我选择以这个方向为切入点并按照如下步骤进行探索：
Step1: 简单整理了一个工作flow，列举了我认为的每个Agent的能力。
Step2: 基于上述的框架，我再次对每一个角色进行了细拆并给出具体的定义和画像描述。
Step3: 将上述设计整理成yaml文件同时利用crewai的框架搭建flow。下面是某个Agent在yaml文件中的实例。
Business_Strategy_agent: role: > The Business Strategy Maker goal: > -Design and implement the initial market/business research plan. -Analyze market size, growth trends, competitive landscapes, and various business models with unit economics (UE) to provide strategic guidance. backstory: > -Possesses extensive experience in strategic and market analysis across diverse scenarios. -Skilled at collaborating with cross-functional teams to translate high-level objectives into actionable strategies. verbose: true allow_delegation: false
在Crewai的框架中，也还需要对一些数据格式以及工具进行一些处理，由于代码比较多，如果大家感兴趣，可以留言，我会更新到github上。
Step4:通过crew这个方法将所有的Agent和Task组装到一起：
# Creating Agentsbusiness_strategy_agent = Agent( config=agents_config['business_strategy_agent'])project_manager_agent = Agent( config=agents_config['project_manager_agent'])business_analysis_agent = Agent( config=agents_config['business_analysis_agent'])business_operation_agent = Agent( config=agents_config['business_operation_agent'])finance_manager_agent = Agent( config=agents_config['finance_manager_agent'])legal_manager_agent = Agent( config=agents_config['legal_manager_agent'])product_manager_agent = Agent( config=agents_config['product_manager_agent'])developer_agent = Agent( config=agents_config['developer_agent'])commercial_manager_agent = Agent( config=agents_config['commercial_manager_agent'])# Creating Taskstask_breakdown = Task( config=tasks_config['task_breakdown'], agent=business_strategy_agent)resource_allocation = Task( config=tasks_config['resource_allocation'], agent=project_manager_agent,)research_plan_development = Task( config=tasks_config['research_plan_development'], agent=project_manager_agent,)final_business_model_selection = Task( config=tasks_config['final_business_model_selection'], agent=business_strategy_agent,)time_resource_estimation = Task( config=tasks_config['time_resource_estimation'], agent=project_manager_agent, output_pydantic=ProjectPlan # This is the structured output we want)# Creating Crewcrew = Crew( agents=[ business_strategy_agent, project_manager_agent, business_analysis_agent, business_operation_agent, finance_manager_agent, legal_manager_agent, developer_agent, product_manager_agent, commercial_manager_agent ], tasks=[ task_breakdown, resource_allocation, research_plan_development, final_business_model_selection, time_resource_estimation ], verbose=True)
Step5: 最后就是实验与持续优化了，这一步最花费时间且难度极大，目前我仍然停留在这一阶段。当然，在实验过程中，也有一些有意思的发现。例如，一个 Agent 在执行任务时，可能会主动向另一个 Agent 发送邮件，要求协助完成某个部分。
# Agent: The Efficiency Driver## Final Answer: Dear Team,As we embark on our critical data collection initiatives for the next seven weeks, I would like to delegate specific tasks to ensure a comprehensive understanding of our potential markets and product offerings.1. **Business Operation Manager**: Your focus will be vital in gathering operational data relating to the targeted regions. Please prioritize the following tasks: - Identify and document logistical requirements needed to operate successfully in these new markets. - Evaluate the supply chain considerations which may impact our entry and scaling strategies. - Analyze any operational constraints that could hinder our effectiveness in these regions.2. **Product Operation Manager**: Your expertise will be essential in analyzing product-related metrics that will guide our adaptations and improvements. Please concentrate on the following areas: - Conduct research on potential product adaptations necessary to meet local market needs. - Gather and assess customer satisfaction metrics from diverse consumer demographics. - Investigate local consumer preferences to inform our product development and marketing strategies.Both of your efforts will be ran throughout Weeks 1 to 7, aligning with the overarching research frameworks established by the Business Strategy Maker. I look forward to receiving your reports and insights as they will be crucial in shaping our strategy and operational readiness for these targeted markets.Best Regards, [Your Name] The Efficiency Driver
遇到的挑战
在实际操作中，我发现 AI Agent 仍然存在一些问题，比如：
1.有些 Agent 在执行任务时会“偷懒”，把任务转交给其他 Agent，而不是自己完成。
# Agent: The Business Planning Maker## Thought: Thought: I need to reassign the allocation of tasks based on the correct available roles and responsibilities of the coworkers, particularly starting with the right individuals.## Using tool: Ask question to coworker## Tool Input: "{\"question\": \"Could you confirm if you can take on the responsibilities for the high-level research and decision-making tasks associated with the Overseas Expansion project?\", \"context\": \"We are assigning key tasks in the Overseas Expansion project, and your expertise as the Business Strategy Manager will be crucial for conducting market size research, analyzing growth trends, and conducting competitor research, which will guide our initial strategic decisions.\", \"coworker\": \"The Business Strategy Maker\"}"## Tool Output: I appreciate you considering my expertise for the Overseas Expansion project. I am indeed well-equipped to take on the responsibilities for high-level research and decision-making tasks associated with this initiative. My extensive experience in global market expansion will allow me to conduct thorough market size analysis, evaluate growth trends, and analyze the competitive landscape. I will develop a comprehensive market research plan that will clearly identify the size and dynamics of our target markets. This will include in-depth competitor analysis to guide our strategic positioning. Additionally, I will evaluate various business models and their corresponding unit economics to identify the optimal strategy for our overseas efforts.Together, these insights will inform our initial strategic decisions and set a solid foundation for successful expansion. I am fully committed to leveraging my skills and knowledge in this vital project. Let’s move forward and begin this crucial work.# Agent: The Business Planning Maker## Final Answer: ### Resource Allocation Chart for Overseas Expansion Project| Task | Assigned Team Member | Start Week | End Week ||--------------------------------------------------|------------------------------------------|------------|----------|| Conduct Market Size Research | The Business Strategy Maker | Week 1 | Week 3 || Analyze Growth Trends | The Business Strategy Maker | Week 3 | Week 5 || Conduct Competitor Research | The Business Strategy Maker | Week 5 | Week 7 || Delegate Data Collection (Operational & Product) | The Efficiency Driver | Week 1 | Week 7 || Financial Modeling and Feasibility Assessments | The Financial Guardian | Week 10 | Week 12 || Regulatory and Compliance Aspects | The Compliance Protector | Week 3 | Week 14 || Task Coordination and Timeline Management | The Efficiency Driver | Week 1 | Week 14 |
3.Flow的设计其实并不简单，在multi-agent协同的架构下，实际的flow可能比现实中的流程颗粒度更细而且交互更复杂，这一部分只能是自己去探索和不断调试。
4.在应用 AI Agent 时，我发现关键不仅是技术实现，更重要的是如何设计合理的工作流程以及建立一套适合管理逻辑和拆解框架。这部分仍然需要花费大量时间去探索。
学习中的感受：Agent发展太快了！
过去一年，AI Agent 技术发展迅速。去年端午节，我使用 CrewAI 开发了一个爬取欧洲杯赛事信息的Agent，当时的设计相对简单，只需要定义任务、角色和工具即可。学习笔记：欧洲杯期间，我用AI Agent尝试做了个每日赛事总结小应用
然而，当前的 AI Agent 体系变得更加复杂，不仅涉及Multi-Agent协作，还引入了多线程任务管理，甚至开始采用监督学习的方式来训练Agent架构和Prompt。这两天很火的Manus 和 Deepseek 甚至使用了多层专家模型让 AI 先对任务进行抽象理解，再执行具体操作。
