Context Management
*Earlier in this chapter, [How Coding Agents Work](how-agents-work.md) showed you the agent loop: read context → plan → act → observe → repeat. The quality of that loop depends almost entirely on the "read context" step. Get context right, and agents produce remarkable work. Get it wrong, and no…
上下文管理
*本章开头的[编程智能体工作原理](how-agents-work.md)向你展示了智能体循环:读取上下文 → 规划 → 行动 → 观察 → 重复。这个循环的质量几乎完全取决于"读取上下文"这一步。上下文做对了,智能体能产出出色的工作;做错了,再多的提示词调整也无济于事。* [How Long Contexts Fail](https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html) 超长上下文窗口并不能自动使智能体变得更聪明——如果不加以管理,额外的上下文会成为负担而非资产。
How Coding Agents Work
*In [Chapter 1](../01-prompt/README.md), you learned to communicate with coding agents — giving them context, writing clear prompts, and building persistent memory systems. Now, let's look under the hood: what actually happens when you press Enter?* *From simple autocomplete to autonomous digital…
编程智能体的工作原理
*在 [第一章](../01-prompt/README.md) 中,你学习了如何与编程智能体沟通——向它们提供上下文、编写清晰的提示词,以及构建持久化的记忆系统。现在,让我们掀开引擎盖看看底层:当你按下回车键时,究竟发生了什么?* *从简单的代码补全到自主运转的数字工程师——深入剖析当今顶尖 AI 编程工具背后的系统架构、上下文策略与核心设计模式。*
Human-Agent Collaboration Modes
Coding agents do not remove engineering judgment. They amplify it. The practical skill is knowing when to explore with an agent, when to turn intent into a spec, when to let the agent execute, and when to slow down for verification. The old question was whether you should "vibe code" or "spec…
人机协作模式:从探索到可验证交付
编程智能体不会取消工程判断。它会放大工程判断。真正重要的能力,是知道什么时候用智能体探索,什么时候把意图固化成规格,什么时候放手让它执行,以及什么时候必须慢下来验证。 过去的问题是:应该 vibe coding,还是 spec-first coding?这个问题曾经有用,但对真实工程来说太窄。