Back to Details

agentic-workflow-guide

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Agentic Workflow Guide

Agent工作流指南

Design, review, and improve agent workflows based on proven principles.
基于经验证的原则，设计、审核并优化Agent工作流。

When to Use

适用场景

ActionTriggers
CreateNew 
.agent.md
, workflow architecture, scaffolding
ReviewOrchestrator not delegating, design principle check, context overflow
UpdateAdding Handoffs, improving delegation, tool configuration
操作触发条件
创建新建
.agent.md
文件、工作流架构搭建、基础框架生成
审核编排器未执行委托、设计原则检查、上下文溢出
更新添加任务交接、优化委托机制、工具配置

Core Principles

核心原则

→ references/design-principles.md
TierPrinciples
EssentialSSOT, SRP, Simplicity First, Fail Fast, Feedback Loop
QualityTransparency, Gate/Checkpoint, DRY, Observability
ScaleHuman-in-the-Loop, Loose Coupling, Graceful Degradation
See design-principles.md > Simplicity First for Anthropic's key recommendation.
→ references/design-principles.md
层级原则
基础必备SSOT, SRP, 简洁优先, 快速失败, 反馈循环
质量保障透明度, 关卡/检查点, DRY, 可观测性
规模化人在回路, 松耦合, 优雅降级
请查看design-principles.md > 简洁优先，这是Anthropic的核心建议。

Workflow Patterns

工作流模式

→ references/workflow-patterns/overview.md
PatternWhen to Use
Prompt ChainingSequential tasks with validation
RoutingProcessing varies by input type
ParallelizationIndependent tasks run together
Orchestrator-WorkersDynamic task decomposition
Evaluator-OptimizerRepeat until quality criteria met
Stop Conditions (MANDATORY): Define success/failure criteria and exit conditions for every loop.
→ references/workflow-patterns/overview.md
模式适用场景
Prompt链式调用带验证的串行任务
路由模式处理逻辑随输入类型变化
并行化独立任务并行执行
编排器-工作者动态任务分解
评估器-优化器重复执行直至满足质量标准
终止条件（强制要求）： 为每个循环定义成功/失败标准及退出条件。

Design Workflow

设计流程

  1. Requirements - Goal, subtasks, dependencies, quality criteria
  2. Pattern Selection - [!] Ask user to confirm pattern before proceeding
  3. Design Diagram - Visualize with Mermaid
  4. Principle Check - Validate against review checklist
  5. Implement & Iterate - Build small → verify → improve
  1. 需求分析 - 目标、子任务、依赖关系、质量标准
  2. 模式选择 - [!] 在继续前请确认用户选定的模式
  3. 设计图 - 使用Mermaid可视化
  4. 原则校验 - 对照审核清单验证
  5. 实现与迭代 - 小步构建 → 验证 → 优化

When to Escalate

升级触发条件

→ references/splitting-criteria.md
LevelConfigurationEscalation Triggers
L0Single PromptRetry 3+, unstable output
L1Prompt + InstructionsSteps > 5, "missed/overlooked" errors
L2Single AgentMultiple responsibilities, context > 70%
L3Multi-AgentIndependent subtasks needed
Quick Check: Prompt > 50 lines? Steps > 5? SRP violation? Context > 70%? → Consider splitting.
→ references/splitting-criteria.md
级别配置类型升级触发条件
L0单一Prompt重试3次以上、输出不稳定
L1Prompt + 操作指南步骤超过5个、出现"遗漏/忽略"类错误
L2单一Agent承担多项职责、上下文占用率超过70%
L3多Agent需要独立子任务
快速检查： Prompt超过50行？步骤超过5个？违反SRP？上下文占用率超过70%？→ 考虑拆分。

Review Checklist

审核清单

→ references/review-checklist.md
  • Single responsibility per agent? (SRP)
  • Errors detected immediately? (Fail Fast)
  • Small iterative steps? (Iterative)
  • Results verifiable at each step? (Feedback Loop)
→ references/review-checklist.md
  • 每个Agent是否单一职责？（SRP）
  • 是否能立即检测到错误？（快速失败）
  • 是否采用小步迭代？（迭代式）
  • 每个步骤的结果是否可验证？（反馈循环）

Key References

核心参考资料

TopicReference
Prompt Templatereferences/prompt-template.md
agentreferences/agent-guide.md
Agent Templatereferences/agent-template.md
Context Managementreferences/context-engineering.md
Handoffsreferences/handoffs-guide.md
Scaffold Toolreferences/scaffold-usage.md
Deep Agentreferences/deep-agent-patterns.md
Agent Evaluationreferences/agent-evaluation.md
主题参考链接
Prompt模板references/prompt-template.md
Agent指南references/agent-guide.md
Agent模板references/agent-template.md
上下文管理references/context-engineering.md
任务交接references/handoffs-guide.md
脚手架工具references/scaffold-usage.md
深度Agentreferences/deep-agent-patterns.md
Agent评估references/agent-evaluation.md

agent Quick Fix

Agent快速修复

Problem: Orchestrator says "I'll delegate" but does work directly.
Solution: Use MUST/MANDATORY language. See agent-guide.md.
yaml
undefined
问题： 编排器表示"我将委托任务"但实际自行处理工作。
解决方案： 使用MUST/MANDATORY这类强制语气。请查看agent-guide.md。
yaml
undefined

MANDATORY: Sub-agent Delegation

MANDATORY: Sub-agent Delegation

You MUST use agent for each file. Do NOT read files directly.
undefined
You MUST use agent for each file. Do NOT read files directly.
undefined

Tools Reference

工具参考

→ references/agent-template.md
PurposeVS Code CopilotClaude Code
Shell
runInTerminal
Bash
Read
read/readFile
Read
Edit
edit/editFiles
Write
Subagent
agent
Task
Web fetch
web/fetch
(MCP)
→ references/agent-template.md
用途VS Code CopilotClaude Code
终端执行
runInTerminal
Bash
读取文件
read/readFile
Read
编辑文件
edit/editFiles
Write
子Agent调用
agent
Task
网络请求
web/fetch
(MCP)

External References

外部参考资料

Official Documentation

官方文档

Design Principles

设计原则

Instructions & Context

操作指南与上下文

Community Resources

社区资源

Prompt Engineering

Prompt工程

Done Criteria

完成标准

  • Workflow pattern selected and confirmed with user
  • .agent.md
    file created with clear Role/Workflow/Done Criteria
  • Design principles checklist passed
  • Agent registered in AGENTS.md (if applicable)
  • 已选定工作流模式并与用户确认
  • 已创建
    .agent.md
    文件，包含清晰的角色/工作流/完成标准
  • 通过设计原则审核清单
  • 已在AGENTS.md中注册Agent（如适用）