
Ars Technica 总结了代码代理从业者的观察:上下文策略已经成为真正的杠杆——而不是底层模型。
用简单的语言来说。 Ars Technica (2026/07/20) 发表了对 Vinay Perneti (Augment Code) 的采访,他主张超越代码编写 AI 框架中的 grep 搜索,转而采用结构化上下文(语义索引器、符号导航器、本地内存)。这不是一个工具的辩论——这正是区分能够快速编写代码的代理和能够生成代码的代理的关键。
自从 2024-25 年编码代理的爆发(Cursor、Cline、Aider、Codex、Claude Code)以来,标准协议仍然非常原始:grep -R foo → 将匹配的文件传递给模型 → 希望。这对于局部任务有效;但一旦涉及跨调用流程的错误、横向约定或推断类型,就会出现问题。最先进的框架已经转向混合模式:专用子代理、通过嵌入的语义回调、通过类型调用的导航(LSP、tree-sitter)。
框架成为战略单位,而不仅仅是模型。没有上下文的前沿模型,其性能不如在良好框架中的较为普通的模型。这颠覆了 IT 部门的采购分析:在上下文堆栈(索引器、语义、导航器、内存)上进行裁决比在底层模型上更具影响力。推论:模型之间的可移植性体现在框架层面,而非 API 层面。KEEL CRUX 观察:在 IDE 使用中,此类堆栈的延迟开销,在内部测试中测量,仍然可被吸收;这不是限制因素。
三个组件构成了现代上下文:
所有这些由一个路由器控制,决定每个任务调用哪个组件。
本文由人工智能撰写,并经人工编辑审核。
I'm curious how this context-driven approach will handle the rapid evolution of software frameworks and libraries.
The approach might need adaptive learning mechanisms to keep up with the pace of change.
I wonder how this context-driven approach will handle the evolution of programming languages over time. Will it adapt or become obsolete?
Interesting question! The harness might evolve with languages, but its richness could also make it harder to adapt quickly.
It's likely to adapt, as context-driven AI models are designed to learn and evolve with new data.
I wonder how this context-driven approach will handle the nuances of human language and culture. It's a complex challenge.
I agree, context is key. But how do we ensure the context is accurate and unbiased?
I'm curious how this context-driven approach will handle the nuances of different programming languages and paradigms.
I've noticed this trend too. Context is indeed becoming more crucial than the underlying model.
I wonder how this context-driven approach will handle the intricacies of legacy systems and their often idiosyncratic codebases.
Interesting point. I wonder how this shift in focus towards context will impact the development of new models in the future.
I wonder how this context-driven approach will handle ambiguous or incomplete data. It's a fascinating shift, but not without challenges.
Harness Ops : post-mortems et bench des agents en prod