📄 中文摘要
软件开发正从“编写代码”转向“策划意图”。借助现代大型语言模型（LLMs），AI代理能够生成大量实现代码，但前提是代理必须扎根于项目的真实背景。AI代理可以被视为具有顺行性失忆症的开发者，能够进行推理和编写代码，但如果没有记忆，它无法可靠地“记住”系统。在以AI为中心的工作流程中，这种记忆由两个主要的构件构成：源上下文和提示历史。源上下文是定义需求、约束、架构和不变性的模块化文档，而提示历史则是捕捉决策、反馈和理由的持续对话。
AI代理、源上下文与提示历史：一种新的软件开发范式
📄 English Summary
AI Agents, Source Context, and Prompt History: A New Software Development Paradigm
Software development is transitioning from 'writing code' to 'curating intent.' With modern large language models (LLMs), a significant portion of implementation can be generated by an AI agent, provided that the agent is grounded in the project's truth. An AI agent can be likened to a developer with anterograde amnesia; it can reason and write code, but it does not reliably 'remember' the system unless given memory. In an AI-first workflow, this memory is constructed from two primary artifacts: source context and prompt history. Source context refers to the curated, modular documentation that defines requirements, constraints, architecture, and invariants, while prompt history captures the ongoing dialogue of decisions, feedback, and rationale.