How-toSelected

10 principles for multi-agent R&D practices

Why we picked it

For developers who use AI for coding but still face bottlenecks from manual intervention, this article explains how to redesign workflows to let agents handle more tasks.

Kiro team released an engineering article titled 'Frontier Engineering' by Clare Liguori, outlining 10 principles for multi-agent R&D. The core idea is that software development practices are dividing: some teams just replace coding tools, while pioneers are trying to restructure workflows, shifting from writing every line of code to building infrastructure for agents to autonomously close loops, aiming for significant or step-like delivery speedups.

Image · 小互
Original · 小互

从换工具到改写工作方式: 面向多 Agent 研⁠发实践的10 条原则 许多人仍把自己放在 AI 与终端之间,充当实时的人肉路由器、异常传递员和点击确认器。 Kiro 团队发布了由 Clare Liguori 主笔的工程专题《Frontier Engineering》(前沿工程学),系统梳理了 10 条面向多代理(Agent)研发实践的探索性原则。 核心主张在于:当前的软件研发实践正在出现分化。 一部分团队只是在既有流程中换了编码工具,而前沿探索者则在尝试重构工作方式——从亲自手写每一行代码,转向搭建让 Agent 自主闭环推进的工程基础设施,以追求显著乃至阶跃式的交付提速。 如果你已经熟悉 AI 辅助编程,却发现端到端交付依旧频繁受阻于人肉卡点,本文将结合 Kiro 的 10 条原则,解析这套工作流背后的运行机制、依赖条件与工程边界。 10 条原则阅读地图 Architect, not typist → 角色重置:从敲击代码的打字员,转向定义意图的架构师 Maximize agent time → 摆脱人肉卡点:把「60 秒等待循环」拉长为「30 分钟自愈闭环」 Build for agents → 改造代码库:像对待「每次都在重新入职的新员工」一样准备环境 Build a fast feedback loop → 快速反馈闭环:从本地校验工具到基于属性的测试补充 Direction over execution → 角色重置中的方向与规格驱动 Treat code as disposable → 代码可重写,边界测试是契约 Hold the line on human standards → 人类标准与审查分工 Trust the boundaries → 防护栏工程 Use agents for everything → 全流程协同 Constantly tune your setup → 持续调优的飞轮 best.xiaohu.ai/article/fronti… 💬 0 🔄 0 ❤️ 2 👀 1075 📊 1 ⚡