agent-workflow

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Agent Workflow Designer

Agent工作流设计器

Overview

概述

This skill guides the design and architecture of AI agent workflows using proven methodologies. When a user presents a problem, this skill helps structure an agent-based solution following the 9-step building process and 8-layer architecture framework validated at Meta.
本技能基于验证过的方法论,指导AI Agent工作流的设计与架构。当用户提出问题时,本技能会遵循Meta验证的9步构建流程和8层架构框架,帮助构建基于Agent的解决方案。

Workflow Decision Tree

工作流决策树

When a user shares a problem or requests agent design help:
  1. Assess the problem scope
    • Is the problem clearly defined? → Proceed to Problem Analysis
    • Is the problem vague? → Ask clarifying questions about desired outcomes and constraints
  2. Determine architecture complexity
    • Simple task (single action)? → Single agent with basic tools
    • Complex task (multiple sub-tasks)? → Consider multi-agent orchestration
    • Integration task (connecting systems)? → Focus on Layer 4 (Tooling) design
  3. Follow the appropriate workflow
    • New agent from scratch → Apply 9-Step Building Process
    • Existing agent improvement → Focus on specific layers needing enhancement
    • Tool integration problem → Apply MCP and tooling patterns
当用户分享问题或请求Agent设计帮助时:
  1. 评估问题范围
    • 问题是否定义清晰?→ 进入问题分析环节
    • 问题是否模糊?→ 询问关于预期结果和约束条件的澄清问题
  2. 确定架构复杂度
    • 简单任务(单一动作)?→ 配备基础工具的单Agent
    • 复杂任务(多个子任务)?→ 考虑多Agent编排
    • 集成任务(连接系统)?→ 重点设计第4层(工具层)
  3. 选择合适的工作流
    • 从零构建新Agent → 应用9步构建流程
    • 优化现有Agent → 聚焦需要改进的特定层级
    • 工具集成问题 → 应用MCP和工具集成模式

9-Step Agent Building Process

9步Agent构建流程

Use this sequential workflow when designing a new agent from scratch:
从零设计新Agent时,使用以下顺序工作流:

Step 1: Define Purpose and Scope

步骤1:定义目标与范围

Key principle: Start with job-to-be-done, not technology.
Ask the user:
  • What specific outcome does the end user need?
  • What are the constraints (budget, time, resources)?
  • What's the success metric?
Bad scope example: "An AI assistant for customer service"
Good scope example: "An agent that takes customer complaints, pulls order history from Shopify API, and drafts refund approvals for orders under $200"
Decision point: Narrow scope = better performance. Resist building Swiss Army knives.
核心原则: 从需完成的任务出发,而非技术。
询问用户:
  • 终端用户需要达成什么具体成果?
  • 有哪些约束条件(预算、时间、资源)?
  • 成功衡量标准是什么?
反面示例: "面向客户服务的AI助手"
正面示例: "一款接收客户投诉、从Shopify API调取订单历史、为200美元以下订单起草退款审批的Agent"
决策要点: 范围越窄,性能越好。避免打造"瑞士军刀"式的通用Agent。

Step 2: Structure Inputs and Outputs

步骤2:结构化输入与输出

Treat the agent as a function with structured interfaces:
Inputs:
  • Use JSON schemas or Pydantic models, not free text
  • Define required vs. optional fields
  • Specify data types and validation rules
Outputs:
  • Return data objects, not prose
  • Define clear error states
  • Include confidence scores when relevant
Example structure:
json
Input: {
  "complaint_text": "string",
  "customer_id": "string",
  "order_id": "string (optional)"
}

Output: {
  "action": "approve_refund | escalate | request_info",
  "refund_amount": "number",
  "reasoning": "string",
  "confidence": "number"
}
将Agent视为具备结构化接口的函数:
输入:
  • 使用JSON schema或Pydantic模型,而非自由文本
  • 定义必填与可选字段
  • 指定数据类型和验证规则
输出:
  • 返回数据对象,而非散文式文本
  • 定义清晰的错误状态
  • 相关场景下包含置信度评分
示例结构:
json
Input: {
  "complaint_text": "string",
  "customer_id": "string",
  "order_id": "string (optional)"
}

Output: {
  "action": "approve_refund | escalate | request_info",
  "refund_amount": "number",
  "reasoning": "string",
  "confidence": "number"
}

Step 3: Write System Instructions

步骤3:编写系统指令

Critical: Spend 80% of design time here.
Include in system prompt:
  • Role definition: "You are a sales qualification specialist..."
  • Behavioral guidelines: "Always ask for budget before proposing solutions"
  • Output format requirements: Specify JSON structure, word limits, tone
  • Edge case handling: What to do when data is missing or ambiguous
Testing strategy: A great system prompt can make GPT-3.5 outperform poorly prompted GPT-4.
关键提示: 投入80%的设计时间在此环节。
系统提示需包含:
  • 角色定义: "你是一名销售资质审核专员..."
  • 行为准则: "在提出解决方案前,始终询问预算情况"
  • 输出格式要求: 指定JSON结构、字数限制、语气
  • 边缘场景处理: 数据缺失或模糊时的应对方式
测试策略: 优秀的系统提示可让GPT-3.5的表现优于提示不佳的GPT-4。

Step 4: Enable Reasoning and External Actions

步骤4:启用推理与外部动作

ReAct Framework Pattern:
  1. Reason: Analyze the current state and decide next action
  2. Act: Call an API, use a tool, or make a decision
  3. Observe: Review the result and determine if goal is achieved
Start simple:
  • Begin with if/then logic before complex reasoning chains
  • Add tools incrementally (don't overwhelm with 50 tools at once)
  • Test each tool integration independently
Common tools to integrate:
  • Calculators for math operations
  • Web browsers for research
  • Database queries for data retrieval
  • API calls to external systems
ReAct框架模式:
  1. 推理: 分析当前状态,决定下一步动作
  2. 执行: 调用API、使用工具或做出决策
  3. 观察: 审查结果,判断是否达成目标
从简开始:
  • 先使用if/then逻辑,再引入复杂推理链
  • 逐步添加工具(不要一开始就集成50个工具)
  • 独立测试每个工具的集成效果
常见集成工具:
  • 用于数学运算的计算器
  • 用于调研的网页浏览器
  • 用于数据检索的数据库查询
  • 调用外部系统的API

Step 5: Orchestrate Multiple Agents (When Needed)

步骤5:多Agent编排(按需使用)

When to use multi-agent architecture:
  • Task has clearly separable sub-tasks
  • Different sub-tasks require different expertise
  • Parallel processing would improve speed
When NOT to use multi-agent:
  • Simple linear workflows
  • Tasks that require continuous context
  • When handoff complexity exceeds benefit
Common 4-agent pattern:
  1. Research Agent: Gathers information from sources
  2. Analysis Agent: Processes and synthesizes data
  3. Writing Agent: Creates structured outputs
  4. QA Agent: Reviews quality and accuracy
Keep handoffs simple: Complex orchestration = complex failures.
何时使用多Agent架构:
  • 任务包含可清晰拆分的子任务
  • 不同子任务需要不同专业能力
  • 并行处理可提升速度
何时不使用多Agent:
  • 简单线性工作流
  • 需要持续上下文的任务
  • 切换复杂度超过收益的场景
常见4-Agent模式:
  1. 调研Agent: 从各类来源收集信息
  2. 分析Agent: 处理并整合数据
  3. 撰写Agent: 创建结构化输出内容
  4. 质检Agent: 审查内容质量与准确性
简化切换流程: 复杂编排意味着复杂故障。

Step 6: Implement Memory and Context

步骤6:实现记忆与上下文管理

Three types of memory to consider:
Conversation history:
  • What happened this session
  • Recent user interactions
  • Current task state
User context:
  • User preferences and settings
  • Past interaction patterns
  • Historical decisions
Knowledge retrieval:
  • Relevant information from knowledge base
  • Similar past cases
  • Domain-specific context
Implementation guidance:
  • Start with simple conversation buffers
  • Add vector databases only when needing semantic search across large datasets
  • Consider memory retrieval latency in architecture
需考虑三类记忆:
对话历史:
  • 当前会话发生的内容
  • 近期用户交互记录
  • 当前任务状态
用户上下文:
  • 用户偏好与设置
  • 过往交互模式
  • 历史决策记录
知识检索:
  • 知识库中的相关信息
  • 类似过往案例
  • 领域特定上下文
实施指导:
  • 从简单的对话缓冲区开始
  • 仅在需要对大型数据集进行语义搜索时,再添加向量数据库
  • 架构设计中需考虑记忆检索延迟

Step 7: Add Multimedia Capabilities

步骤7:添加多媒体能力

Modern agents should handle:
  • Voice input/output for accessibility
  • Image understanding for visual tasks
  • Document processing (PDF, DOCX, spreadsheets)
Strategic approach: Add capabilities based on actual user needs, not "nice-to-haves."
现代Agent应支持:
  • 用于无障碍访问的语音输入/输出
  • 用于视觉任务的图像理解
  • 文档处理(PDF、DOCX、电子表格)
策略建议: 根据实际用户需求添加能力,而非"锦上添花"的功能。

Step 8: Format and Deliver Results

步骤8:格式化与交付结果

Output is your product's UX. Design outputs for:
Human consumption:
  • Clear formatting and structure
  • Scannable with headers and bullets
  • Professional appearance
System consumption:
  • Valid JSON/XML
  • Consistent field names
  • Error codes for handling
Quality standard: Great agent outputs look like a human created them.
输出即产品的用户体验。 设计输出时需兼顾:
人类阅读:
  • 清晰的格式与结构
  • 可快速浏览的标题与项目符号
  • 专业的呈现形式
系统调用:
  • 合法的JSON/XML格式
  • 一致的字段名称
  • 用于处理的错误码
质量标准: 优秀的Agent输出应看起来如同人类创作。

Step 9: Build Interface or API

步骤9:构建界面或API

Delivery method options:
  • Chat interface for conversational tasks
  • API endpoints for system integration
  • Integration with existing tools (Slack, email, CRM)
Best practice: The best agents feel invisible—they just make things happen.
交付方式选项:
  • 适用于对话任务的聊天界面
  • 用于系统集成的API端点
  • 与现有工具(Slack、邮件、CRM)集成
最佳实践: 优秀的Agent应"无形"——默默完成任务即可。

8-Layer Architecture Framework

8层架构框架

When analyzing agent architecture needs, consider which layers require attention:
分析Agent架构需求时,需明确哪些层级需要重点关注:

Layer 1: Infrastructure

第1层:基础设施

Foundation: Cloud, databases, APIs, compute resources
Key considerations:
  • GPU/TPU requirements for inference
  • Data storage and retrieval speed
  • Load balancing for scale
  • Monitoring and observability
Common mistake: Underestimating compute needs—agents make more API calls than traditional apps.
基础: 云服务、数据库、API、计算资源
关键考虑因素:
  • 推理所需的GPU/TPU
  • 数据存储与检索速度
  • 用于扩容的负载均衡
  • 监控与可观测性
常见误区: 低估计算需求——Agent的API调用量远超传统应用。

Layer 2: Agent Internet

第2层:Agent网络

Operating system for agents: Identity, state management, inter-agent communication
Current state: Mostly custom-built, but platforms like LangChain and CrewAI are emerging.
Agent的操作系统: 身份管理、状态管理、Agent间通信
当前现状: 大多为定制开发,但LangChain、CrewAI等平台正在兴起。

Layer 3: Protocol

第3层:协议

Standards for interoperability: MCP (Model Context Protocol) is becoming the standard
Key principle: Bet on open standards, not proprietary solutions. MCP allows any tool to work with any agent.
互操作性标准: MCP(Model Context Protocol)正成为通用标准
核心原则: 押注开放标准,而非专有解决方案。MCP可让任意工具与任意Agent协作。

Layer 4: Tooling Enrichment

第4层:工具增强

Agent superpowers: RAG systems, function calling, external integrations
Quality over quantity: 5 rock-solid tools > 50 flaky integrations
Tool categories:
  • Data retrieval (databases, APIs)
  • Computation (calculators, processors)
  • Communication (email, messaging)
  • Content creation (documents, reports)
Agent的超能力: RAG系统、函数调用、外部集成
质量优先: 5个可靠工具 > 50个不稳定集成
工具类别:
  • 数据检索(数据库、API)
  • 计算(计算器、处理器)
  • 通信(邮件、消息工具)
  • 内容创作(文档、报告)

Layer 5: Cognition Reasoning

第5层:认知推理

The brain: Planning, decision-making, error handling
Critical elements:
  • Guardrails to prevent hallucinations
  • Error recovery strategies
  • Confidence scoring
  • Graceful degradation
User forgiveness: Users forgive agents that fail gracefully, not ones that spiral into nonsense.
核心大脑: 规划、决策、错误处理
关键要素:
  • 防止幻觉的防护机制
  • 错误恢复策略
  • 置信度评分
  • 优雅降级
用户容错性: 用户会原谅优雅失败的Agent,而非陷入混乱的Agent。

Layer 6: Memory Personalization

第6层:记忆个性化

Human touch: Personal context, preferences, conversation history
Start simple: Store user preferences and conversation context before building complex personalization.
人文关怀: 用户个人上下文、偏好、对话历史
从简开始: 先存储用户偏好与对话上下文,再构建复杂个性化功能。

Layer 7: Application

第7层:应用层

User-facing products: The actual agent functionality users interact with
Focus strategy: Nail one use case before expanding to others.
用户面向的产品: 用户实际交互的Agent功能
聚焦策略: 先做好一个用例,再扩展其他场景。

Layer 8: Ops Governance

第8层:运维治理

Risk management: Monitoring, cost control, privacy, oversight
Build from day one: Retrofitting governance is expensive and painful.
Key components:
  • Cost tracking per agent action
  • Privacy enforcement and data handling
  • Human-in-the-loop for critical decisions
  • Audit logs and compliance
风险管理: 监控、成本控制、隐私、监督
从第一天开始构建: 事后补充治理成本高昂且痛苦。
关键组件:
  • 按Agent动作跟踪成本
  • 隐私合规与数据处理
  • 关键决策的人工介入
  • 审计日志与合规性

Problem-to-Solution Workflow

问题到解决方案的工作流

When a user presents a problem:
Step 1: Clarify the problem
  • What's the current manual process?
  • What's the desired outcome?
  • What are the constraints (time, cost, technical)?
  • What data sources are available?
Step 2: Assess agent appropriateness Not every problem needs an agent. Consider:
  • Is the task repetitive and rule-based?
  • Does it require decision-making with context?
  • Would automation provide significant value?
  • Is the problem scope clear and bounded?
Step 3: Map to architecture Using the 8 layers, identify which need focus:
  • Simple task → Focus on Layers 4, 5, 7 (tools, reasoning, application)
  • Complex integration → Add Layer 3 (protocol) emphasis
  • Scalability concern → Prioritize Layers 1, 8 (infrastructure, ops)
Step 4: Design workflow Apply the 9-step building process, calling out:
  • Critical decision points
  • Tool integration requirements
  • Multi-agent needs (if any)
  • Memory and context strategy
Step 5: Identify implementation path Based on user's role and resources:
  • For PMs: High-level architecture and tool selection
  • For engineers: Detailed technical implementation with code patterns
  • For product teams: Full stack from requirements to monitoring
当用户提出问题时:
步骤1:澄清问题
  • 当前手动流程是怎样的?
  • 预期成果是什么?
  • 有哪些约束条件(时间、成本、技术)?
  • 可用的数据来源有哪些?
步骤2:评估Agent适用性 并非所有问题都需要Agent。需考虑:
  • 任务是否重复且基于规则?
  • 是否需要结合上下文做决策?
  • 自动化能否带来显著价值?
  • 问题范围是否清晰且明确?
步骤3:映射到架构 利用8层架构,确定需重点关注的层级:
  • 简单任务 → 聚焦第4、5、7层(工具、推理、应用)
  • 复杂集成 → 加强第3层(协议)的设计
  • 扩容需求 → 优先关注第1、8层(基础设施、运维)
步骤4:设计工作流 应用9步构建流程,明确:
  • 关键决策点
  • 工具集成需求
  • 多Agent需求(如有)
  • 记忆与上下文策略
步骤5:确定实施路径 根据用户角色与资源:
  • 产品经理: 高层架构与工具选型
  • 工程师: 带代码模式的详细技术实现
  • 产品团队: 从需求到监控的全流程方案

Tool Integration Patterns

工具集成模式

MCP (Model Context Protocol) Integration

MCP(模型上下文协议)集成

When tools support MCP:
  1. Agent discovers available tools
  2. Agent calls tools using standardized interface
  3. Tool returns structured response
  4. Agent processes and continues workflow
Advantage: Write once, use with any agent.
当工具支持MCP时:
  1. Agent发现可用工具
  2. Agent通过标准化接口调用工具
  3. 工具返回结构化响应
  4. Agent处理响应并继续工作流
优势: 一次编写,可与任意Agent协作。

Custom API Integration

自定义API集成

When building custom integrations:
  1. Define clear API contract (inputs/outputs)
  2. Implement error handling and retries
  3. Add rate limiting and caching
  4. Monitor usage and costs
  5. Document for agent consumption
构建自定义集成时:
  1. 定义清晰的API契约(输入/输出)
  2. 实现错误处理与重试机制
  3. 添加限流与缓存
  4. 监控使用情况与成本
  5. 编写供Agent调用的文档

Common Integration Scenarios

常见集成场景

CRM Integration (Salesforce, HubSpot):
  • Read customer data
  • Create/update records
  • Search across objects
  • Trigger workflows
Communication Tools (Slack, Email):
  • Send messages/notifications
  • Read incoming requests
  • Monitor channels
  • Respond to mentions
Data Sources (Databases, APIs):
  • Query structured data
  • Retrieve documents
  • Search knowledge bases
  • Aggregate information
CRM集成(Salesforce、HubSpot):
  • 读取客户数据
  • 创建/更新记录
  • 跨对象搜索
  • 触发工作流
通信工具集成(Slack、邮件):
  • 发送消息/通知
  • 读取 incoming 请求
  • 监控频道
  • 回复提及消息
数据源集成(数据库、API):
  • 查询结构化数据
  • 检索文档
  • 搜索知识库
  • 聚合信息

Decision Framework: Single vs. Multi-Agent

决策框架:单Agent vs 多Agent

Use Single Agent When:

适用单Agent的场景:

  • Task is linear and sequential
  • Context must be maintained throughout
  • Decision-making is unified
  • Complexity of orchestration > benefit
  • 线性顺序任务
  • 全程需保持上下文
  • 决策需统一
  • 编排复杂度超过收益

Use Multi-Agent When:

适用多Agent的场景:

  • Clear task separation exists
  • Sub-tasks need different expertise
  • Parallel processing improves performance
  • Quality benefits from specialization
Example - Customer Support:
Single agent sufficient for: "Take customer complaint, pull order history, draft refund approval"
Multi-agent beneficial for: "Monitor social media, categorize issues, research solutions, generate responses, escalate critical cases, track resolution"
  • 任务可清晰拆分
  • 子任务需要不同专业能力
  • 并行处理可提升性能
  • 专业化可提升质量
示例 - 客户支持:
单Agent即可胜任: "接收客户投诉、调取订单历史、起草退款审批"
多Agent更具优势: "监控社交媒体、分类问题、调研解决方案、生成回复、升级关键案例、跟踪解决进度"

Common Pitfalls and Solutions

常见陷阱与解决方案

Pitfall 1: Scope Creep

陷阱1:范围蔓延

Problem: Trying to build a general-purpose assistant Solution: Define narrow, specific job-to-be-done with clear success metrics
问题: 试图构建通用型助手 解决方案: 定义狭窄、具体的任务目标,明确成功衡量标准

Pitfall 2: Tool Overload

陷阱2:工具过载

Problem: Giving agent 50+ tools upfront Solution: Start with 5 essential tools, add incrementally based on actual needs
问题: 一次性给Agent配备50+工具 解决方案: 从5个核心工具开始,根据实际需求逐步添加

Pitfall 3: Skipping System Prompt

陷阱3:忽略系统提示

Problem: Generic or minimal instructions Solution: Invest 80% of time crafting detailed system prompt with examples and edge cases
问题: 使用通用或极简指令 解决方案: 投入80%的时间编写包含示例与边缘场景的详细系统提示

Pitfall 4: No Error Handling

陷阱4:无错误处理

Problem: Agent breaks on unexpected inputs Solution: Design graceful degradation, clear error states, and fallback behaviors
问题: Agent在意外输入下崩溃 解决方案: 设计优雅降级、清晰错误状态与 fallback 行为

Pitfall 5: Ignoring Costs

陷阱5:忽略成本

Problem: Runaway API costs from inefficient agent design Solution: Build cost monitoring from day one, implement caching, optimize prompt length
问题: 低效的Agent设计导致API成本失控 解决方案: 从第一天开始构建成本监控,实现缓存、优化提示长度

Pitfall 6: Over-Engineering Architecture

陷阱6:过度架构设计

Problem: Building all 8 layers simultaneously Solution: Start with Layers 4, 5, 7 (tools, reasoning, application), add others as needed
问题: 同时构建全部8层架构 解决方案: 从第4、5、7层(工具、推理、应用)开始,再按需添加其他层级

Output Format

输出格式

When providing agent workflow solutions, structure the response as:
  1. Problem Restatement: Confirm understanding of the user's need
  2. Agent Architecture Recommendation: Single vs. multi-agent, with rationale
  3. Step-by-Step Workflow: Apply relevant steps from the 9-step process
  4. Tool Integration Plan: Specific tools needed and integration approach
  5. Layer Analysis: Which of the 8 layers need focus and why
  6. Implementation Guidance: Prioritized next steps based on user's role
  7. Success Metrics: How to measure if the agent is working
提供Agent工作流解决方案时,按以下结构组织响应:
  1. 问题重述: 确认对用户需求的理解
  2. Agent架构建议: 单Agent或多Agent选型及理由
  3. 分步工作流: 应用9步流程中的相关步骤
  4. 工具集成计划: 所需具体工具与集成方式
  5. 层级分析: 需重点关注的8层架构及原因
  6. 实施指导: 根据用户角色确定优先下一步动作
  7. 成功指标: 如何衡量Agent的工作效果

Agent Taxonomy Quick Reference

Agent分类速查

When users ask about existing tools:
Category 1: Consumer Agents (Built-In)
  • Examples: ChatGPT Agent, Claude, Gemini, Grok
  • Best for: Quick tasks, research, content creation
  • User type: Everyone, especially PMs
Category 2: No-Code Builders
  • Examples: Zapier Central, n8n, Make
  • Best for: Workflow automation without coding
  • User type: PMs, operations teams
Category 3: Developer-First Platforms
  • Examples: LangChain, CrewAI, AutoGen, Swarm
  • Best for: Custom agent features in products
  • User type: Engineering teams
Category 4: Specialized Agent Apps
  • Examples: Cursor (coding), Perplexity (research), Notion AI (writing)
  • Best for: Specific job-to-be-done with deep specialization
  • User type: Domain-specific professionals
当用户询问现有工具时:
类别1:内置消费级Agent
  • 示例:ChatGPT Agent、Claude、Gemini、Grok
  • 最佳场景:快速任务、调研、内容创作
  • 用户群体:所有人,尤其是产品经理
类别2:无代码构建平台
  • 示例:Zapier Central、n8n、Make
  • 最佳场景:无需编码的工作流自动化
  • 用户群体:产品经理、运营团队
类别3:开发者优先平台
  • 示例:LangChain、CrewAI、AutoGen、Swarm
  • 最佳场景:在产品中构建自定义Agent功能
  • 用户群体:工程团队
类别4:专业Agent应用
  • 示例:Cursor(编码)、Perplexity(调研)、Notion AI(写作)
  • 最佳场景:具备深度专业化的特定任务
  • 用户群体:领域专业人士",