AI-SDLC Design Phase: From Spec to Parallel Sub-Agent Tasks
A published AI-SDLC practice describes turning a specification, decision record, and fixed code baseline into a human-reviewed technical plan and a task graph. Tasks are organized around independently verifiable business behaviors, then executed by parallel sub-agents whose context is limited to their assigned task, required files, write scope, and completion criteria. The workflow also defines how missing materials pause only affected tasks and how developers confirm the plan and development scope.
The article provides a concrete workflow for coordinating isolated coding agents, including behavior-based task decomposition, human approval, dependency scheduling, and write-scope controls.
需求阶段交出来的是一份 spec、一份决定记录和一个固定的代码基线。设计阶段拿着这三样东西做两件事:写出技术方案,再把方案拆成一组开发任务。
开发阶段由多个子 agent 并行执行这些任务。每个子 agent 只看得到派给它的那个任务,看不到设计阶段的对话,也看不到别的任务。所以设计阶段的目标就一条:开发时不需要猜。
本篇讲六件事:
- 设计阶段读什么、交出什么;
- 架构问题怎么问,Agent 想多做的东西怎么拦住;
- 技术方案文档怎么写,为什么只生成一次、由人来审;
- 任务怎么拆,每个任务要写清什么,以及 Claude Code 和 Codex 的子 agent 启动时能看到什么;
- 任务说明怎么编译成任务图,开发时怎么调度;
- 缺资料时怎么办,方案和开发范围怎么确认。
其中一处是我们踩过的坑:早期在设计阶段内跑独立评审和多轮自审,阶段内一出现循环就非常耗时,后来改成方案只生成一次、由人来审。