Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Code-map MCP tools cut tokens for coding-agent localization tasks

AI summary

A University of York study, as reported by Juejin AI, compared five code representations—AST, call graph, dependency graph, documentation, and a hybrid view—exposed through MCP tools against a search-only baseline on a fixed Claude Code stack. For localization tasks, the reported configurations saved 25–39% of tokens while improving recall, but gains were limited when the bottleneck was generating correct patches or tests; a call-graph configuration also improved success on SecBench.

In-depth insight

The study suggests that code-map tools are most useful when coding agents need to locate relevant code, rather than serving as a general-purpose route to higher task success.

氛围图

💡 一句话总结:约克大学把 5 种代码表示(AST/调用图/依赖图/文档/混合)做成 MCP 工具,在固定 Claude Code 栈上与纯搜索基线跑了 4 个仓库级基准的受控对比。结论:代码地图是省钱工具不是提分工具——定位是瓶颈的任务 token 省 25–39%、召回还在涨;瓶颈在"写对补丁/测试"的任务基本白挂。做 agent 基建的人值得读完。

导语:你的 agent token 都烧在哪了?

先问一个可能戳到你的问题:你有没有盯着 Claude Code(或任何 coding agent)的用量面板,看着它为一个其实不复杂的 issue 反复 grep、read、再 grep、再 read,最后烧掉几十万 token 才摸到那个该改的文件?

这种"探索税"几乎是所有仓库级 agent 的宿命。主流 agent(Claude Code、SWE-agent、Agentless、OpenHands)都把仓库当纯文本:靠词法搜索和 shell 命令一层层扒。"谁定义了这个符号?谁调用它?哪些模块依赖它?"——这些结构信息明明就在仓库里,agent 却要用一串昂贵的试错把它们重新拼出来。

#University of York#Claude Code#Haiku 4.5#MCP#LinuxFLBench#SWE-bench Pro#SecBench#SWT-bench

Original · 10 min readRead the original
Group chat

No comments yet. Start the conversation.