Django 联合创始人 Simon Willison 探讨 AI 编程代理如何跨越临界点，让开发者 95%的代码无需手动输入，同时警告随着不安全使用增加，即将发生'挑战者号灾难'。
Simon Wilson, co-creator of Django, discusses how AI coding agents have crossed a threshold, enabling developers to produce 95% of code without typing, while warning of an impending 'Challenger disaster' as unsafe usage grows.
要点 · TL;DR
2025 年 11 月，编程智能体跨越门槛，代码基本按指令运行。 Coding agents crossed a threshold in Nov 2025, making code mostly work as instructed.
瓶颈从编写代码转向构思、测试和质量保证。 The bottleneck shifts from writing code to ideation, testing, and quality assurance.
提示注入仍未解决；AI 安全可能发生‘挑战者号灾难’。 Prompt injection remains unsolved; a 'Challenger disaster' in AI safety is possible.
核心观点 · Key points
2025年11月编码智能体跨越了门槛：代码基本按指令工作，使可靠的智能体式工程成为可能。 Coding agents crossed a threshold in Nov 2025: code mostly works as instructed, enabling reliable agentic engineering.
编写代码现在成本低廉；瓶颈转移到构思、原型设计、测试和质量保证。 Writing code is now cheap; the bottleneck shifts to ideation, prototyping, testing, and quality assurance.
测试驱动开发（红/绿TDD）对编码智能体至关重要，以确保代码运行并保持质量。 Test-driven development (red/green TDD) is crucial for coding agents to ensure code runs and maintains quality.
AI放大资深工程师的能力，但威胁到缺乏深厚专业知识来利用智能体的中层职业人士。 AI amplifies experienced engineers but threatens mid-career professionals who lack deep expertise to leverage agents.
提示注入仍然是一个未解决的安全问题；致命三重奏（私人数据、恶意输入、泄露）是危险的。 Prompt injection remains an unsolved security problem; the lethal trifecta (private data, malicious input, exfiltration) is dangerous.
AI安全中的偏差常态化可能导致‘挑战者号灾难’——一场本可避免的重大灾难。 The normalization of deviance in AI safety may lead to a 'Challenger disaster'—a major preventable catastrophe.
反共识 · Contrarian takes
用好编码智能体在精神上令人疲惫；资深工程师因并行管理多个智能体而精疲力竭。 Using coding agents well is mentally exhausting; experienced engineers burn out by managing multiple agents in parallel.
沉迷AI的人比以往更努力；生产力提升带来更多雄心，而非闲暇。 AI pilled people work harder than ever; productivity gains lead to more ambition, not leisure.
手工编写的人类代码正成为优质资产；数据标注公司购买2022年之前的GitHub仓库。 Handcrafted human-written code is becoming a premium asset; data labeling companies buy pre-2022 GitHub repos.
最佳编码模型与它们用SVG绘制骑自行车鹈鹕的能力相关——无人知晓原因。 The best coding models correlate with how well they draw a pelican riding a bicycle in SVG—no one knows why.
Open Claw成功是因为Anthropic和OpenAI没有构建它；他们害怕其拥抱的安全风险。 Open Claw succeeded because Anthropic and OpenAI didn't build it; they feared security risks it embraces.
记者特别适合使用AI，因为他们将其视为不可靠的来源，就像人类线人一样。 Journalists are uniquely suited to use AI because they treat it as an unreliable source, like human informants.
本期章节 · Chapters（共 35）
编码代理与11月转折点Coding agents and the November inflection point
代码是AI代理最易解决的问题Code as the easiest problem for AI agents
赞助商消息：WorkOSSponsor message: WorkOS
氛围编码及其局限Vibe coding and its limits
专业人员的代理工程Agentic engineering for professionals
无压力写作Writing without pressure
前沿：用编码代理构建专业软件Frontier: building professional software with coding agents
暗工厂的作用What the dark factory does
氛围编码的QA代理与工厂概念Vibe-coded QA agents and the factory concept
AI从构建扩展到QA与构思AI expanding from building to QA and ideation
AI原型时代的人类价值Human value in the age of AI prototyping
AI作为头脑风暴伙伴AI as brainstorming companion
AI对职业中期专业人士的影响AI impact on mid-career professionals
AI热潮与期望带来的疲惫Exhaustion from AI boom and expectations
人们低估了变化的速度People underestimate how quickly things change
代理工程模式手册The agentic engineering pattern book
AI堆栈与工具AI Stack and Tools
搜索集成取代谷歌Search integration replacing Google
鹈鹕骑自行车基准Pelican riding bicycle benchmark
SVG中的代码注释与中文模型Code comments in SVG and Chinese models
秘密替代方案与实验室作弊Secret alternatives and labs cheating
AI转型中的奇思与乐趣Whimsy and joy in AI transition
鹈鹕骑自行车Pelican on a bicycle
使用LLM组合工具Using LLMs to Combine Tools
红绿测试驱动开发Red-Green Test-Driven Development
AI生成的测试与代码质量AI-generated tests and code quality
提示注入与致命三重奏Prompt injection and the lethal trifecta
提示注入与致命三重奏Prompt Injection and the Lethal Trifecta
预测记录与火鸡类比Prediction track record and the turkey analogy