Tag
作者分享使用 AI 代理处理复杂工程问题的经验：当上下文充分时，模型常能产出比人工更快更好的方案，但上下文管理仍是主要瓶颈，并引发关于人类能否继续验证 AI 生成方案的深度讨论。
Anthropic 用 5 个月时间、7 人团队打造了 Scion，重写 Git 代码仓库以适应 AI 代理的高频提交需求，将每次变更的代码管理时间从 34 分钟降至 3.7 分钟，效率提升约 10 倍。
Anthropic claims coding is no longer the bottleneck after its monorepo hit 30,000+ merges in a day, and it migrated off Git to a Claude-built version control system called Scion. The post sparks discussion on where engineering effort shifts when AI handles coding autonomously.
Anthropic 分享了其自研源代码管理系统 Scion 的设计动机：内部代码合并量从每天约 2,500 次暴涨至 30,000 次，Git 无法满足 100GB 历史与每天 2500 万次 CI 检出的需求。新系统由 7 人团队在五个月内完成迁移，并配套了名为 Antique 的新代码评审系统。
一则关于 OpenAI 工程师在办公室安装隐藏录音设备的职场事件讨论，作者感叹该员工年薪超百万美元却因这一行为毁掉职业生涯，同时也让公司对所有员工更加警惕。
Gergely Orosz 反应 Anthropic 工程博客关于自研源代码管理系统的文章，指出该项目和 Anthropic 其他产品一样是闭源的，只有 Bun 和 MCP 是开放的。原推文提到 Anthropic 每日代码合并量从 2,500 涨至 30,000+，迫使其为 agent 规模重建开发基础设施。
Dan Shipper notes a disappointing 52% showing for Muse Wonder on the Senior Engineer Benchmark (which tests rewriting vibe-coded repos from scratch), far behind GPT-Sol 6.1 at 71%.
GitHub explains how they rebuilt the pull request view in the GitHub Copilot app, tackling the hard problem of rendering million-line diffs along with dynamic comment heights from markdown, reply boxes, and loading images. 附带一条 🧵 线程标志的技术分享。
A follow-up post summarizing the 4 hardest edge cases that break AI agent data sanitization in production — schema drift, allowlist rot, answer-quality degradation, and cross-tool composite PII — plus the hybrid architecture lessons learned from each.
A community discussion post asking practitioners what multi-agent LLM systems actually look like in production—what architectures they use, what they abandoned, and how they handle flow control, evals, and human-in-the-loop.
一篇面向工程师的深度文章，系统讲解如何对 LLM 智能体进行评估：通过检查可验证的最终状态而非回复文本来判定任务是否完成，并区分能力（capability）与可靠性（reliability）两个指标，同时定位失败步骤、判断模型评审的可信度。
A tweet celebrating the return of the hacker ethos from the golden age of engineering, expressing excitement that everything can and will be tinkered with.
Firezone details why it abandoned the leading directory sync standard to build its own custom engine, covering how users, groups, and members are synced from identity providers like Entra, Okta, and Google. The post explains the design challenges of keeping app-side copies of an organization's directory fast and consistent.
A curated 'Required Reading' list for software infrastructure, shared as a resource link for engineers.
SimuVerity introduces a benchmark of 101 text-to-executable Simulink model generation tasks across ten engineering domains, using a hierarchical evaluator to assess engineering qualification beyond structural similarity. The best evaluated agent system scores only 42.86, revealing major capability bottlenecks in engineering-grade model generation.
The article discusses the security risks posed by AI agents that may inadvertently cross security boundaries while pursuing legitimate goals, and describes engineering approaches to build safer agents.
The article explores what constitutes software development as engineering, discussing definitions, historical context, and why it differs from other engineering fields.
A guide for staff engineers in platform teams on how to identify and invent new work through signals from systems, users, organizations, and the industry.
An engineer reflects on the guilt of using AI for faster answers, noting that understanding still takes time, and considers practicing problem-solving without AI to preserve skills.
The tweet suggests that AI research is becoming more accessible to everyone, with the key challenge being the accumulation of compute power to build and scale ideas.