Reddit 用户分享 Claude 多智能体开发流程
一名 Reddit 用户分享了一套多智能体 Claude 开发流程,由所有者、编排器、构建器、审查器和合并器分工处理任务。流程还包含按任务类型切换模型、失败升级、上下文交接、敏感操作审批及发布前测试。用户称它多数时候有效,但交接期间有时会忽略所有者规则,也没有数据证明其能提升效率或减少 token 使用。
该帖具体展示了一套在临近上线时结合多个 Claude 智能体与人工审批的开发流程,也记录了上下文交接可能破坏规则执行的失败情况。
I am trying to determine if this is a good way to structure Claude and improve efficiencies and token usage. It is kinder working *Most of the time* the issue is when handovers take place sometimes the Orch goes off the rails and ignore owner rules. It just tends to do its own thing.
I have limited the Orch levels to a maximum of 8, thereafter I start the Orch again manually.
Context handover: Windows are 250k = Wind down = 280k start preparing for handover.
Failure Guard:
Sonnet First on Low/Med and tasks - First fail move to Opus - Second fail dead stop.
Opus on High/Schema and tasks - First fail move to Fable Review/Audit - Second test Opus with fable guideline - Third fail dead stop - log for redesign.