📄 中文摘要
在与具备代理能力的集成开发环境（IDE）交互时，如 antigravity，当用户请求解释或生成代码时，实际发生的过程并非简单的提示传递。IDE 会构建一个综合输入，包括文件、代码、选择和搜索等上下文信息，而不仅仅是用户的提示。整个工作流程从用户的请求开始，经过上下文构建、语言模型（LLM）的预测、工具调用（如有必要）、执行层的操作，最终返回结果并再次由 LLM 决定下一步动作。这一过程确保了 IDE 能够准确返回用户所需的内容。
AI IDE 的实际工作原理 - 深入剖析
出处: How AI IDEs Actually Work - Under the Hood
发布: 2026年3月22日
📄 English Summary
How AI IDEs Actually Work - Under the Hood
When interacting with an agentic IDE like antigravity, the process triggered by user requests for explanations or code generation involves more than just sending a prompt. The IDE constructs a combined input that includes context such as files, code, selections, and searches, rather than solely relying on the user's prompt. The overall workflow begins with the user request, followed by context building, language model (LLM) predictions, tool calls if necessary, execution layer actions, and finally returning results, which are then used by the LLM to decide the next steps. This process ensures that the IDE can accurately deliver the requested content.