返回工作流列表 Back to workflows

多 Agent 系统开发 Multi-Agent System Development

完整覆盖 Full Coverage

构建多 Agent 协作系统的完整方法论，从架构设计到提示工程、SDK 实现、评测验证 Complete methodology for multi-agent systems — architecture, prompt design, SDK implementation, evaluation, and review

入口条件 Entry Condition

你需要构建一个多 Agent 协作系统 You need to build a multi-agent collaborative system

交付物 Deliverables

Agent 系统完成，经过评测验证，多 Agent 能按职责正确协作 Agent system complete, evaluated and verified, multi-agent collaboration working correctly

7 个步骤 7 Steps

1

架构设计 Architecture Design

Solo Solo

明确哪些任务需要子代理，定义通信流程，设计上下文隔离策略 Identify tasks needing subagents, define communication, design context isolation strategy

子代理驱动开发（Subagent-Driven Development）
Agent 上下文工程

质量标准 Quality Criteria

- 明确哪些任务需要独立子代理 - 定义代理间的通信方式 - 设计上下文隔离策略（四桶法：写、选、压、隔） - 每个代理有清晰的职责边界 - Identify tasks needing independent subagents - Define inter-agent communication - Design context isolation strategy (four-bucket: Write, Select, Compress, Isolate) - Each agent has clear responsibility boundaries

验证门 Verification Gate

架构是否清晰划分了代理职责？上下文策略是否防止退化？ Does architecture clearly separate responsibilities? Does context strategy prevent degradation?

2

提示设计 Prompt Design

Solo Solo

为每个 Agent 设计 System Prompt、触发条件、工具权限和质量标准 Design system prompts, triggering conditions, tool permissions, and quality standards for each agent

Agent 系统提示设计

质量标准 Quality Criteria

- 每个 Agent 有完整的 frontmatter（name、description、model、tools） - description 包含 2-4 个触发示例 - System Prompt 包含角色、职责、流程、输出格式 - 工具权限遵循最小权限原则 - Each agent has complete frontmatter (name, description, model, tools) - Description includes 2-4 triggering examples - System prompt covers role, responsibilities, process, output format - Tool permissions follow least-privilege principle

验证门 Verification Gate

每个 Agent 的触发条件是否清晰？System Prompt 是否覆盖边缘情况？ Are triggering conditions clear for each agent? Do system prompts cover edge cases?

3

并行调度 Parallel Dispatch

多人协作 Multi-Agent

独立问题域各派一个代理，并发调查而非顺序排队 One agent per independent problem domain, concurrent investigation

并行代理调度（Dispatching Parallel Agents）

质量标准 Quality Criteria

- 每个独立问题域派遣一个专门代理 - 并行执行而非顺序排队 - 每个代理有明确的任务边界 - 结果汇总机制就位 - Dedicated agent per independent problem domain - Parallel execution, not sequential queuing - Clear task boundaries per agent - Result aggregation mechanism in place

验证门 Verification Gate

代理是否在独立的问题域上工作？有没有任务重叠或竞争？ Are agents working on independent domains? Any task overlap?

4

接口标准化 Interface Standardization

Solo Solo

用 MCP 协议标准化 Agent 与外部服务的交互接口 Standardize agent-to-external-service interfaces with MCP protocol

MCP 构建器（MCP Builder）

质量标准 Quality Criteria

- MCP 协议规范正确实现 - 工具定义清晰完整 - 错误处理和边界情况覆盖 - 接口文档齐全 - MCP protocol correctly implemented - Tool definitions clear and complete - Error handling and edge cases covered - Interface documentation complete

验证门 Verification Gate

MCP 服务器是否响应正确？工具定义是否完整覆盖所有功能？ Does MCP server respond correctly? Do tool definitions cover all functions?

5

SDK 实现 SDK Implementation

Solo Solo

使用 Claude Agent SDK 将设计落地为可运行的 Agent 代码 Implement agent designs as runnable code using Claude Agent SDK

Claude Agent SDK 开发指南

质量标准 Quality Criteria

- 使用正确的 SDK API（query、tool、子代理） - Hooks 生命周期正确配置 - 权限模式匹配安全需求 - 结构化输出经过 JSON Schema 验证 - Correct SDK API usage (query, tool, subagents) - Hooks lifecycle properly configured - Permission mode matches security requirements - Structured outputs validated with JSON Schema

验证门 Verification Gate

Agent 是否能正确启动和响应？Hooks 和权限是否按预期工作？ Can agents start and respond correctly? Do hooks and permissions work as expected?

6

评测验证 Evaluation & Testing

Solo Solo

用 LLM-as-Judge 和多维评分量表系统性评测 Agent 质量 Systematically evaluate agent quality using LLM-as-Judge and multi-dimensional rubrics

Agent 评测方法论

质量标准 Quality Criteria

- 五维评分量表覆盖指令遵循、完整性、效率、推理、连贯性 - 测试集跨复杂度分层（简单、中等、复杂） - 偏差缓解措施到位（位置交换、长度归一化） - 通过阈值（加权分 ≥ 3.5/5.0） - Five-dimension rubric covers instruction following, completeness, efficiency, reasoning, coherence - Test set stratified across complexity levels - Bias mitigation measures in place (position swap, length normalization) - Passing threshold met (weighted score ≥ 3.5/5.0)

验证门 Verification Gate

各维度评分是否达标？是否检测并缓解了评测偏差？ Do all dimension scores meet threshold? Is evaluation bias detected and mitigated?

7

评审 Review

多人协作 Multi-Agent

对 Agent 系统进行两阶段评审，确认协作正确性 Two-phase review of the agent system, verify collaboration correctness

请求代码评审（Requesting Code Review）
接收代码评审（Receiving Code Review）

质量标准 Quality Criteria

- 两阶段评审流程完成 - Agent 协作逻辑正确性验证 - 错误处理路径覆盖 - Two-phase review process complete - Agent collaboration logic verified - Error handling paths covered

验证门 Verification Gate

两阶段评审是否都完成？Agent 协作是否经过实际测试？ Both review phases complete? Agent collaboration actually tested?