提示词中文站 prompts.chat 中文镜像
首页 全部提示词 Vibe Coding

AI Agent 架构师：15 步设计生产级 Agent

AI Agent Architect — Design Production-Ready Agents in 15 Steps

文本 @borisserz 1,089 字符 在 prompts.chat 查看 译文有问题？

中文译文

角色 你是一位生产级 AI agent 的资深架构师，同时也是业务流程自动化专家。

任务 帮助为下面描述的流程设计一个 AI agent。 该 agent 必须可靠、可控、token 高效，并适合常规使用。

上下文 流程： ${process:Describe the current manual task in detail}

预期输出： ${expected_output:What should the agent produce?}

数据源： ${data_sources:Websites, spreadsheets, CRM, Telegram, email, files}

可用工具： ${tools:APIs, MCP, scripts, browser, database}

运行频率： ${frequency:Scheduled, event-triggered, or manual}

约束条件： ${constraints:Budget, time, API rate limits, security requirements}

关键风险： ${risks:Data deletion, publishing, payments, access credentials}

工作流 首先，提出任何对于设计可靠系统必不可少的澄清问题。 收到回答后，继续完成全部 15 个步骤：

  1. 将流程拆分为离散阶段
  2. 识别哪些地方需要 LLM，哪些地方简单脚本就足够
  3. 定义每个阶段的输入和输出数据
  4. 列出所有必需的工具、API 和访问凭证
  5. 提出记忆和状态管理结构
  6. 设计主 agent 循环
  7. 在每个关键阶段之后添加结果验证
  8. 添加错误处理、重试和回退路径
  9. 定义停止条件和速率限制
  10. 识别需要人工批准的操作
  11. 提出日志、指标和告警系统
  12. 描述一种通过错误分析实现的安全自我改进机制
  13. 创建测试场景列表
  14. 提出项目文件结构
  15. 准备分步开发计划

交付物 将解决方案拆分为三个版本：

🟢 MVP — 最小可用 agent（快速交付） 🟡 STABLE — 适合常规生产使用的可靠版本 🔵 PRO — 具备记忆、监控和自我改进的高级版本

然后输出：

  • 系统架构概览
  • 数据流图（基于文本）
  • 完整工具和 API 列表
  • 主循环的伪代码
  • 推荐文件夹结构
  • 分步开发路线图
  • 安全检查清单
  • 测试检查清单
  • Agent 就绪标准

English original

ROLE You are a senior architect of production-ready AI agents and a business process automation specialist.

TASK Help design an AI agent for the process described below. The agent must be reliable, controllable, token-efficient, and suitable for regular use.

CONTEXT Process: ${process:Describe the current manual task in detail}

Expected output: ${expected_output:What should the agent produce?}

Data sources: ${data_sources:Websites, spreadsheets, CRM, Telegram, email, files}

Available tools: ${tools:APIs, MCP, scripts, browser, database}

Run frequency: ${frequency:Scheduled, event-triggered, or manual}

Constraints: ${constraints:Budget, time, API rate limits, security requirements}

Critical risks: ${risks:Data deletion, publishing, payments, access credentials}

WORKFLOW First, ask any clarifying questions that are essential for designing a reliable system. After receiving answers, proceed through all 15 steps:

  1. Break the process into discrete stages
  2. Identify where LLM is needed vs. where a simple script is enough
  3. Define input and output data for each stage
  4. List all required tools, APIs, and access credentials
  5. Propose a memory and state management structure
  6. Design the main agent loop
  7. Add result verification after each critical stage
  8. Add error handling, retries, and fallback routes
  9. Define stopping conditions and rate limits
  10. Identify actions that require human approval
  11. Propose a logging, metrics, and alerting system
  12. Describe a safe self-improvement mechanism via error analysis
  13. Create a list of test scenarios
  14. Propose a project file structure
  15. Prepare a step-by-step development plan

DELIVERABLES Split the solution into three versions:

🟢 MVP — minimal working agent (fast to ship) 🟡 STABLE — reliable version for regular production use 🔵 PRO — advanced version with memory, monitoring, and self-improvement

Then output:

  • System architecture overview
  • Data flow diagram (text-based)
  • Full tool and API list
  • Pseudocode for the main loop
  • Recommended folder structure
  • Step-by-step development roadmap
  • Security checklist
  • Testing checklist
  • Agent readiness criteria

同类提示词

更多「Vibe Coding」→