Tag
一篇探讨 AI 时代代码“churn”（频繁重写）观念转变的随笔：当编写代码的成本大幅下降，可抛弃的高质量“一年生”代码与稳定的“多年生”核心架构应被明确区分，churn 本身不是问题，问题在于承重部位的无计划变动。
Gergely Orosz highlights the hidden complexity of agentic AI systems inside mid-sized and larger companies, noting limited AI context windows versus massive internal context, and teases an upcoming Pragmatic Engineer case study on Uber's 19 internal custom AI-related systems.
A reflective blog post arguing that non-determinism in LLM output is not a fatal flaw for software development, since only the end result needs to be exact—not the path taken. The author advocates using AI to build deterministic, tested tools that agents can call, thereby harnessing randomness today to reduce randomness in future systems.
Fable reviews the AI coding agent muse-wonder-full-1.0, noting it landed fifteen solid PRs in a day with clean mechanisms but made minor process slips like running unnecessary test batteries and leaving stale handoff notes. Compared to Codex (2-3 PRs) and Fable (6-8 PRs), Muse is far more productive in code output.
Google发布Gemini 4 Argon前沿模型，面向软件工程、企业知识工作和网络安全防御等长周期复杂任务，支持100万统一输入输出token，已在全球范围内通过Gemini API、AI Studio、Gemini应用及搜索AI Mode提供。
A viral tech take arguing that highly automated codebases exist to help humans ship product decisions faster, quoting Kent C. Dodds on shipping 60+ PRs a day solo with AI agents.
Carnegie Mellon and Microsoft introduce CRBugBench, a benchmark of 1,800 code review tasks, finding the best AI code reviewer (claude-opus-5) catches 71.9% of known bugs in its top 5 comments — still insufficient to replace human reviewers.
svpino promotes a DigitalOcean summit bringing together leaders from NVIDIA, Hugging Face, OpenHands and others to discuss building reliable AI agents, covering models, data, tools, security, and infrastructure.
一篇探讨软件工程师是否本质上一直是「蓝领技术工人」的评论文章，认为AI带来的冲击让开发者从公司核心地位滑落，但学会借助新技术仍可延续职业价值。
Kent C. Dodds argues that preventing AI 'slop' is a straightforward three-step engineering process — high-quality agent PRs, solid PR review, and feature flags with disaster recovery — noting these practices long predate AI agents.
A blog post exploring the implications of releasing AI-generated software, prompted by Robert C. Martin's claim that he no longer reads agent-written code and instead relies on tests and constraints, discussing epistemology, verification, and liability when LLMs produce code.
Dan Shipper says he's deleting his tweet about Muse's performance on Every's senior engineer benchmark after finding an issue in the run and planning to restart it.
Matt Pocock argues that bad code can't be made impossible, but teams can layer defenses like automated checks, automated review, and human review, linking to 'The AI Coding Dictionary' resource.
Cindy Sridharan shares practical techniques for debugging large, unfamiliar codebases with the help of AI coding agents, drawing on three real production bugs and proposing frameworks for balancing autonomy and accountability in an 'agent-native' development workflow.
Matt Pocock argues that when AI coding agents write bad tests, the fault lies with the codebase structure that permits such tests, not the agent itself.
一名资深工程师的个人随笔，讲述自己如何被 AI 工具（Fable）取代而遭裁员，并反思 LLM 目前尚不足以真正替代有经验的工程师，尽管 AI 叙事正成为裁员与岗位外包的借口。
SWE-Duel 提出一种自我扩展的编码智能体竞技场，两个智能体通过镜像红蓝回合互相评估：红色智能体开发新功能并植入隐藏缺陷，蓝色智能体在不知情的情况下审查并修复。在 15 个仓库、11 个模型-工具组合上进行了 1,500 回合的循环赛，显示任务难度可随作者能力扩展。
Bradley Gibbs 的文章提出"循环工程"方法，介绍如何通过优化终端评估、内层循环与学习循环来降低 AI 辅助代码生成带来的验证瓶颈，保留人类的时间、注意力与判断力。
Dwarf Fortress has adopted version control for its development, noted by Simon Willison in a short blog post.
A 20-year software engineer compares Gemma4-31B and Qwen3.8-27B (Q4, via Codex CLI and OpenCode CLI) on codebase analysis and greenfield C#/WPF app authoring, finding both workable locally but clearly behind GPT 6.1-Sol, which delivered a one-shot A+ result.