AI编程工具降低效率的根因与团队管控指南
Why Your AI Coding Tools Are Making Developers Slower
Developers using AI are spending 19% longer on tasks, and slower tools aren’t the reason. Hallucinations invent APIs that don’t exist, security risks multiply from copy-pasted credentials, and senior developers are now reviewing five times more code than before.
使用 AI 的开发者在任务上花费的时间延长了 19%，而工具缓慢并非原因。幻觉会编造不存在的 API，复制粘贴凭证导致安全风险倍增，高级开发人员现在需要审查的代码量是之前的五倍。
At Business of Software USA, Ivan Barajas Vargas, CEO and co-founder of MuukTest, walked through the findings from five of the most cited academic papers on AI in development. The picture: speed and efficiency aren’t the same thing. Code that requires endless review, security audits, and refactoring costs more than it saves.
在 Business of Software USA 大会上，MuukTest 的首席执行官兼联合创始人 Ivan Barajas Vargas 深入剖析了五篇关于开发领域 AI 的高引用学术论文的发现。结论是：速度和效率并不是一回事。需要无尽审查、安全审计和重构的代码，其成本高于它节省的成本。
The problem is how teams are using the tools.
问题在于团队如何使用这些工具。
Tool Overload Hides Real Security Risks
工具过载掩盖了真实的安全风险
There are now hundreds of AI coding tools on the market: Cursor, Copilot, Claude, GitHub’s own offerings, plus specialised tools for every framework. Most teams pick one by gut feel, try a second, and end up running both in production.
市场上现在有数百种 AI 编码工具：Cursor、Copilot、Claude、GitHub 自己的产品，以及针对每个框架的专业化工具。大多数团队凭直觉选择一个，尝试第二个，最终在生产环境中同时运行两者。
That’s shadow AI: unauthorised, unvetted tools loose in the development workflow. A sales intern tries a new tool to “help.” Each one opens its own security surface.
这就是影子 AI：未经授权的、未经验证的工具散落在开发工作流中。一名销售实习生尝试新工具以“提供帮助”。每一个工具都打开了各自的安全攻击面。
The real risk is copy-pasting credentials into a system nobody’s vetted. Once that data leaks, it’s gone.
真正的风险是将凭证复制粘贴到任何人都未验证的系统中。一旦数据泄露，就再也找不回来了。
The fix: pick one stack and enforce it. Write standards. Any team can write a prompt, but the good ones write guidelines for how AI gets used in development, with input from business and product, not just engineering.
解决方案：选择一个技术栈并强制执行。制定标准。任何团队都能编写提示词，但优秀的团队会制定指南，规定如何在开发中使用 AI，并且这些指南需来自业务和产品部门的输入，而不仅仅是工程部门。
“It’s Like Using a Chainsaw”
“就像使用电锯一样”
That’s how Ivan puts it: AI coding tools are powerful in trained hands and a mess in untrained ones. Auto-complete, refactoring, scaffolding, root cause analysis, they all work, provided someone actually knows what they’re doing.
Ivan 就是这样形容的：AI 编码工具在受过训练的人手中威力巨大，而在未经训练的人手中则是一团糟。自动补全、重构、脚手架搭建、根本原因分析，它们都能正常工作，前提是有人真正知道自己在做什么。
AI generates thousands of lines of code, most of it syntactically fine and conceptually plausible. Until you dig in and find APIs that don’t exist. That’s built into how these models work: they extrapolate from training data, and sometimes the extrapolation is fiction.
AI 生成成千上万行代码，其中大部分语法正确且概念上合理。直到你深入挖掘，发现不存在的 API。这是这些模型工作原理的一部分：它们从训练数据中推断，有时这种推断就是虚构的。
That’s tax.
这就是代价。
The fix is context. Load your RAG system or vector database with documentation, architecture decisions, and code standards. Test performance regularly. Don’t assume generated code scales just because it compiles. And keep reviewing outputs: the more your team generates with AI, the more a senior developer ends up validating.
解决方案是提供上下文。将文档、架构决策和代码标准加载到你的 RAG 系统或向量数据库中。定期测试性能。不要仅仅因为生成的代码能编译就假设它能扩展。并且持续审查输出：你的团队用 AI 生成的代码越多，高级开发人员最终需要验证的工作就越多。
When my team ships faster with AI but I’m not sure code review is keeping pace, I want to know whether the tool is the problem or whether we skipped the guardrails.
当我的团队借助 AI 交付得更快时，如果我不确定代码审查是否跟得上节奏，我想知道是工具的问题，还是我们跳过了护栏机制。
The New Job Description: Full-Time Reviewer
新的职位描述：全职审查员
Senior developers are reviewing five times more code than before AI entered the workflow. A junior developer with AI generates more code, faster. But every line needs review, testing, and validation. And the people you actually need for architecture, mentorship, and strategy are stuck validating what the junior team generated.
资深开发人员审查的代码量是 AI 进入工作流前的五倍。一位使用 AI 的初级开发人员能更快、更多地生成代码。但每一行代码都需要审查、测试和验证。而你真正需要用于架构设计、导师指导和战略规划的人员，却被困在验证初级团队生成的代码上。
More code doesn’t mean better code. AI writes like a junior developer: a lot of output, some of it buggy, most of it immature. Tech debt compounds.
更多的代码并不意味着更好的代码。AI 像初级开发人员一样编写代码：产出量大，其中部分存在缺陷，大部分尚不成熟。技术债务随之累积。
The burnout point is the one worth sitting with. A tool that’s supposed to save time is quietly taxing your best people, and most teams haven’t noticed because the dashboard still says “faster.”
倦怠点是值得深入思考的关键所在。一个旨在节省时间的工具正在悄然消耗你最优秀人才的精力，而大多数团队尚未察觉，因为仪表盘上仍然显示“速度更快”。
The fix starts with acknowledging the shift: AI tools multiply senior workload, they don’t eliminate it. Monitor that load. And assign ownership: if someone is responsible for vetting the AI, integrating it, and maintaining standards, the workflow stays manageable. If no one owns it, it metastasises.
解决方案始于承认这一转变：AI 工具放大了资深开发人员的工作量，而非消除它。监控这种负载。并明确责任归属：如果有人负责审核 AI 输出、集成 AI 并维护标准，工作流程就能保持可控。如果无人负责，问题就会恶化蔓延。
Treat AI Like a Chainsaw
将 AI 视为电锯
The teams getting real value from AI aren’t treating it as a replacement for architecture or strategy. They’re treating it like a chainsaw: powerful, but only in trained hands with the guards on.
从 AI 中获得真正价值的团队并未将其视为架构或战略的替代品。他们将其视为电锯：功能强大，但仅限受过训练且装有防护装置的人使用。
That means picking one tool and committing to it. Writing standards. Feeding the system context. Testing for performance. Instrumenting observability from day one, so you know why things break. Naming owners. Watching developer load so your senior people don’t burn out reviewing everyone else’s output.
这意味着选择一个工具并坚持使用它。制定编写规范。为系统提供上下文。进行性能测试。从第一天起就配置可观测性工具，以便了解故障原因。指定负责人。监控开发人员的工作负载，确保资深人员不会因审查他人产出而倦怠。
It also means accepting where vibe coding belongs: prototypes, mockups, small features. Not the core architecture of a fintech platform or a healthcare system. There, AI scaffolds and refactors. It doesn’t decide.
这也意味着接受“氛围编程”（vibe coding）的适用场景：原型、模拟图和小型功能。而非金融科技平台或医疗系统的核心架构。在这些领域，AI 仅负责搭建脚手架和重构代码，并不做决策。
The vibes are only off if you skip the integration work and hope for the best.
只有当你跳过集成工作并寄希望于好运时，感觉才会不对劲。
Watch the Full Talk
观看完整演讲
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力