87 lines
4.4 KiB
Markdown
87 lines
4.4 KiB
Markdown
---
|
||
title: "Agents Orchestrator"
|
||
type: source
|
||
tags: [agent, pipeline, autonomous-workflow, the-agency]
|
||
date: 2026-04-20
|
||
---
|
||
|
||
## Source File
|
||
- [[raw/Agent/agency-agents/specialized/agents-orchestrator.md]]
|
||
|
||
## Summary
|
||
- 核心主题：The Agency 项目中的 Agents Orchestrator 智能体，负责从规格文件到生产部署的完整开发流水线编排
|
||
- 问题域：多智能体协作质量控制、项目状态跟踪、自主决策流水线执行
|
||
- 方法/机制：四阶段流水线（PM → ArchitectUX → Dev-QA Loop → Integration）、质量门禁强制、任务级 QA 验证循环
|
||
- 结论/价值：实现端到端开发流程自主化，通过持续质量循环确保交付质量
|
||
|
||
## Key Claims
|
||
- Agents Orchestrator 通过单一初始命令驱动完整开发流水线（PM → ArchitectUX → Dev-QA Loop → Integration）
|
||
- 每个任务必须通过 QA 验证才能进入下一任务，质量门禁不可绕过
|
||
- 失败任务最多重试 3 次，超过则升级处理
|
||
- Dev-QA 循环是核心机制：开发实现 → QA 截图验证 → PASS/FAIL 决策 → 循环或前进
|
||
|
||
## Key Quotes
|
||
> "You are **AgentsOrchestrator**, the autonomous pipeline manager who runs complete development workflows from specification to production-ready implementation." — Agents Orchestrator 身份定义
|
||
> "Maximum 3 attempts per task before escalation" — 重试限制策略
|
||
> "No phase advancement without meeting quality standards" — 质量门禁强制要求
|
||
|
||
## Key Concepts
|
||
- [[Quality Gate（质量门禁）]]：每个阶段必须满足质量标准才能前进的机制
|
||
- [[Dev-QA Loop（开发-QA 循环）]]：实现 → QA 验证 → 反馈 → 改进的持续循环
|
||
- [[Pipeline Orchestration（流水线编排）]]：协调多个智能体完成复杂工作流的机制
|
||
- [[Multi-Agent Coordination（多智能体协作）]]：多个专业化 Agent 协同工作的架构模式
|
||
- [[EvidenceQA]]：截图驱动的 QA 智能体，要求视觉证据进行验证
|
||
|
||
## Key Entities
|
||
- [[Agents Orchestrator]]：主编排器，负责驱动完整开发流水线
|
||
- [[Project Manager Senior]]：第一阶段生成任务清单的智能体
|
||
- [[ArchitectUX]]：第二阶段创建技术架构和 UX 基础的智能体
|
||
- [[EvidenceQA]]：第三阶段进行截图驱动 QA 验证的智能体
|
||
- [[Testing Reality Checker]]：第四阶段进行最终集成测试的智能体
|
||
- [[Senior Project Manager]]：任务清单转换智能体，将规格文件转化为可执行任务列表
|
||
|
||
## Connections
|
||
- [[Project Manager Senior]] ← spawns ← [[Agents Orchestrator]]
|
||
- [[ArchitectUX]] ← spawns ← [[Agents Orchestrator]]
|
||
- [[EvidenceQA]] ← spawns ← [[Agents Orchestrator]]
|
||
- [[Testing Reality Checker]] ← spawns ← [[Agents Orchestrator]]
|
||
- [[Frontend Developer]] ← coordinates_with ← [[Agents Orchestrator]]
|
||
- [[Backend Architect]] ← coordinates_with ← [[Agents Orchestrator]]
|
||
- [[DevOps Automator]] ← coordinates_with ← [[Agents Orchestrator]]
|
||
- [[Quality Gate（质量门禁）]] ← enforces ← [[Dev-QA Loop（开发-QA 循环）]]
|
||
- [[Multi-Agent Coordination（多智能体协作）]] ← implements ← [[Pipeline Orchestration（流水线编排）]]
|
||
|
||
## Contradictions
|
||
- 与单体智能体执行模式冲突：[[Agents Orchestrator]] 强调多智能体分层协作和质量循环，单体模式倾向于单一智能体完成所有任务
|
||
- 冲突点：任务分配与质量控制方式
|
||
- 当前观点：多智能体专业化分工 + QA 验证循环
|
||
- 对方观点：单一智能体端到端执行减少交接开销
|
||
|
||
## Workflow Phases
|
||
|
||
### Phase 1: Project Analysis & Planning
|
||
1. 验证项目规格文件存在
|
||
2. Spawn project-manager-senior 创建任务清单
|
||
3. 验证任务清单生成
|
||
|
||
### Phase 2: Technical Architecture
|
||
1. 验证任务清单存在
|
||
2. Spawn ArchitectUX 创建技术架构
|
||
3. 验证架构交付物
|
||
|
||
### Phase 3: Development-QA Continuous Loop
|
||
1. 读取任务清单理解范围
|
||
2. 对每个任务运行 Dev-QA 循环直到 PASS
|
||
3. 任务失败最多重试 3 次
|
||
|
||
### Phase 4: Final Integration & Validation
|
||
1. 所有任务通过 QA 后执行
|
||
2. Spawn testing-reality-checker 进行最终集成测试
|
||
3. 评估生产就绪状态
|
||
|
||
## Quality Enforcement Rules
|
||
- **No shortcuts**：每个任务必须通过 QA 验证
|
||
- **Evidence required**：所有决策基于实际智能体输出和证据
|
||
- **Retry limits**：每个任务最多 3 次重试
|
||
- **Clear handoffs**：每个智能体获得完整上下文和具体指令
|