Better tools made Copilot code review worse. Here's how we actually improved it.
GitHub swapped Copilot code review's custom exploration tools for the better-maintained shared CLI tools (grep, glob, view) expecting a clean upgrade — and got higher cost and fewer caught issues. The tools weren't the problem; the instructions were. Rewriting them around how a reviewer actually reads a pull request flipped the regression into ~20% lower review cost at the same quality.
Marked as a signal by the editors; worth tracking.
On this page12 sections
The editors' notes below — why it matters, the context, the learning path — are written in Chinese.
Record
The signal
#When GitHub migrated Copilot code review from its own code exploration tools to the shared Unix-style tools that power Copilot CLI (grep, glob, view), benchmarks showed higher review cost and fewer issues caught — despite the new tools being better maintained. The diagnosis: the instructions, inherited from the old tool layer's design (SWE-agent-style repository navigation), no longer matched the new tools or the task. Rewriting the workflow instructions around how a reviewer actually reads a pull request — anchored on pull request evidence — flipped the regression into roughly 20% lower average review cost while maintaining review quality. The broader lesson: when you build on an agent framework you inherit its tools and their implicit workflows, and as your use case drifts they can quietly start working against you.
Why now
Why it matters
#生产级智能体的负面结果复盘比成功案例稀缺得多。「更好的工具让效果更差」直指一个普遍盲区：工具与指令是耦合的整体，替换其一必须重写另一——而且没有回归基准，这类退化会被安静地上线。
Background
Technical context
#智能体的行为由「工具能力 × 指令假设」共同决定：工具定义能做什么，指令定义怎么探索。继承自框架的工具带着隐含的工作流设计（这里是 SWE-agent 式仓库导航），当任务从「解决 issue」漂移到「审查 PR」，同一套探索模式就从助力变成开销。以任务的原生证据（PR diff）为锚点重写指令，是把通用工具重新对齐到具体任务的正确杠杆。
Audience
Who should care
#Next
Learning path
#- 读原文，注意「换工具前后」的基准对比是怎么设计的
- 对照 Shippy 复盘的「确定性工具层」经验，理解工具层设计的两种取舍
- 为自己的智能体产品建立一个最小回归基准，再做下一次工具或提示词变更
Network
Where it sits in the network
#How this signal connects to neighbouring technologies, skills and background knowledge. Follow any link to keep exploring.
Further thought
Follow-up questions
#- 重写后的审查指令是否会开源或沉淀为可复用模式？
- 20% 成本降低在不同模型上是否稳定成立？
- 同样的「工具×指令错位」诊断方法能否用于其他智能体场景？
Publication record
Featured in
#The briefings that carried this signal, newest first; the last row is the weekly review of its week. The tag is the section it ran in.
技术简报 - 2026-08-09
Worth tracking技术简报 - 2026-07-28
Worth trackingWeekly review · Jul 6 – Jul 12
Worth tracking
Sources