Tag
作者分享使用 DeepSeek V4.1 Flash（EXL3 2.9 bpw，2x DGX Spark，TP2）作为实现代理、配合独立审查代理开发真实功能的实践经验，指出其报告有时会超出实际证据。
SVPino argues companies already have enough model intelligence and instead need better knowledge organization, highlighting Falconer, a tool that links code with a company knowledge base and auto-updates outdated docs on pull request merges.
Entire (cofounded by former GitHub CEO Thomas Tomczak) and Gergely Orosz are hosting a pop-up coffee meetup in San Francisco on October 29 to discuss the future of source control and code review with AI agents.
HyperHat's founder shares results from a benchmark of 60 engineers directing AI coding agents, finding that specification quality, mechanism verification, and review behavior—not prompting tricks—separate strong engineers, and that most people miss partial agent failures.
Stage launches on Product Hunt as a tool for reviewing, fixing, and signing off working prototypes and products, letting teams pin comments on prototypes from Figma, Claude, Cursor, or GitHub and using Claude to auto-fix issues before filing Linear or Jira tickets.
作者回顾一年来AI编写全部代码的经历，分析为何科技行业并未出现预期的AI失业潮，认为高利率才是裁员主因，并提出AI尚未取代岗位的若干原因。
In this GitHub Podcast episode, Senior Product Manager Sameen Karim explains the design motivation behind the native stacked pull requests feature, the architectural trade-offs involved, and the process of rewriting merge logic code that hadn't been touched in ten years to support multi-PR merging.
YC W24 startup Greptile raised a $50M Series B for its AI code review platform, which reviews over 1% of all GitHub pull requests daily and serves thousands of engineering teams including Nvidia and Coinbase.
Whiteboard 是一款开源的 IDE，定位为「为深思熟虑的软件设计而生的画布」，可在统一界面中审查交互式可视化、代码 diff 与 AI 代理轨迹。
ClaudeDevs 披露 Claude 约 90% 的代码由 Claude Code 自身编写，合并 PR 数量翻倍而回退率下降超 40%，仍有约 10% 的代码由 Anthropic 员工手写，Kent C. Dodds 对此表示惊讶。
Anthropic revealed that Claude wrote approximately 90% of the code for Claude Code itself, and shared their experience of doubling the number of merged PRs and reducing the rollback rate by over 40% by improving the code review process.
Imprint shares how its `small-change` approach speeds up pull request review, explaining eligibility criteria and the reasoning behind rethinking review purposes as PR volume and review times grow.
Sphinx 是一个 Git hook 工具，用于 AI 辅助开发：在 agent 生成的代码推送之前，它会基于 diff 提问开发者是否真正理解代码行为，答对才能继续推送。支持 Codex、Claude Code、Cursor 等编码 agent，开源免费。
Claude Code 的开发流程表明，为 AI 代理提供自动化验证机制（如 E2E 测试、属性测试和多代理对抗审查）可让 Claude 自行编写 90% 以上的代码，同时在 PR 合并量翻倍的情况下将回滚率降低 40% 以上。
The article examines how rising AI-generated code is increasing bugs and review load, and surveys automated verification approaches like Superpowers and RoboRev to preserve quality without sacrificing speed.
A Lobsters discussion asking whether teams allow LLMs to write both production code and tests, with the author considering a policy restricting AI generation to only one of the two to avoid tautological, over-engineered results.
Simon Willison 引用了 Addy Osmani、Morgan Westlee Lunt 和 Qing Wang 关于 Anthropic Claude Code 团队代码审查实践的观点，指出 AI 审查让代码审查从发布前的关卡变为写作过程中的持续检查，从而让人类审查专注于更高层次的问题。
Anthropic 分享 Claude Code 团队如何在 AI 写了约 90% 代码的情况下进行代码评审：通过多层自动化检查、把人工注意力集中在高风险环节，并在 PR 数量翻倍的同时将回滚率降低 40% 以上。
作者复盘其共享 markdown 记忆库中的 agent 审批流程漏洞：审批只发生在聊天中、提交滞后或缺失、ingest agent 自行提交了绝大多数 commit，导致『已提交』并不等于『人工审阅过』。他们实施了轻量修复，包括未审阅文件标注、超过 7 天的 open change 周检、以及 review flag 必须指定负责人，并就 review gate 应放在读侧还是写侧征求意见。
Claude Code Review 现已面向 Claude Team 和 Enterprise 计划正式可用，支持 Quick、Standard、Deep 三档审查级别，Boris Cherny 表示它几乎能在 PR 合并前捕获所有 bug，大幅提升开发效率与代码质量。