导读

This article dives into the unexpected side effects of AI coding agents on software development workflows: while code generation becomes faster, engineers are now burdened with more frequent decision-making, heavier code review loads, and rising decision fatigue. The core argument is that the bottleneck of the software development lifecycle has shifted from code writing to human judgment, and we need to rebuild workflow systems to adapt to this new change. The article really made me rethink the so-called "AI productivity improvement" - many times we only see the efficiency of code generation, but ignore the increased pressure on developers in the later stages of the process.

这篇文章深入探讨了AI编程工具给软件开发流程带来的意料之外的副作用：代码生成变快的同时，工程师却要承担更频繁的决策工作、更重的代码审核负担，决策疲劳问题日益凸显。核心观点是软件开发生命周期的瓶颈已经从写代码转移到了人为判断环节，我们需要重构工作流体系来适配这种新变化。读完我才重新理解了所谓的「AI提升效率」——很多时候我们只看到了生成代码的速度，却忽略了开发者在流程后期被增加的压力。

Coding agents are giving everyone decision fatigue

AI编程工具正在让所有人陷入决策疲劳

With much of a software engineer’s time moving from writing code to structuring prompts and reviewing code, the workday is getting denser and more intense. Can AI solve the problems it's causing?

现在软件工程师的大量时间从写代码转向了写提示词和审核代码，工作日的工作密度越来越高、强度越来越大。AI能解决它自己制造的这些问题吗？

{{image_1}}

There’s no doubt that coding agents have changed how software gets built. In the past three years or so, code generators have gone from fancy autocomplete to tools that can whip up a whole application while you wait. Engineers with knowledge of best practices, pitfalls, and the language of software are able to co-create code without having to futz with semicolons and unclosed brackets.

毋庸置疑，AI编程工具已经彻底改变了软件开发的方式。近三年来，代码生成工具已经从高端的自动补全功能，进化到了能在你等待的几分钟里就写出一整个应用的程度。懂最佳实践、了解常见坑、熟悉软件开发语言的工程师，不用再纠结分号漏写、括号没闭合这类琐碎问题，就能和AI协作完成代码创作。

What’s in doubt is whether this change has been productive, cost efficient, or good for developers.

但存疑的是，这种变化到底有没有真的提升生产率、降低成本，或者对开发者更友好。

Easy-to-create code has put greater strain on the later parts of the software development lifecycle (SDLC): code review, DevOps/SRE, security, and infrastructure. It’s also put greater strain on the developers themselves.

生成代码变得简单，反而给软件开发生命周期的下游环节带来了更大压力：代码审核、DevOps/SRE、安全、基础设施环节的负担都变重了。开发者自身的压力也同时飙升。

According to recent data, automation intensity for enterprise users has grown 55% year-over-year, and overall activity has increased 46%. That means the workday hasn’t grown; it’s just gotten denser with work as automations produce more without alleviating the need for humans to decide on what the definition of good is.

最新数据显示，企业用户的自动化强度同比增长了55%，整体工作量增长了46%。这意味着大家的工作时长并没有增加，只是工作密度变高了——自动化产出了更多内容，但并没有减少人类对「什么是好的结果」的判断需求。

{{image_2}}

Code is cheap, code review less so

代码很便宜，代码审核可不便宜

In the pre-AI era, code was expensive because engineers were expensive. They had a ton of knowledge about the languages, processes, and paradigms that produced good software. That led to some pretty lousy metrics: hours spent, lines of code written, commits per day. These were easy to measure and sure looked like a decent approximation of results. Along the way, organizations started looking at outcomes, with schemes like DORA metrics trying to put numbers on those.

前AI时代，代码成本很高，因为工程师的人力成本很高。他们掌握了大量关于编程语言、开发流程、优秀开发范式的知识，才能产出高质量软件。这也催生了很多糟糕的考核指标：工作时长、写的代码行数、每天提交次数。这些指标好统计，看起来也像是能合理反映工作成果。后来企业开始关注实际产出，用DORA指标这类体系试图量化真实的开发效果。

Those lousy metrics are returning with a vengeance, outcomes be damned. Not only are agentic coders bragging about their lines of code stats, but organizations have bragged about their increased developer productivity. Engineers might be writing more code, but they’re reviewing even more. The new tech bro vibes with hustle culture and is AI-pilled.

现在那些糟糕的指标又卷土重来了，没人再关心实际产出。不光用AI编程的开发者炫耀自己的代码行数统计数据，企业也在吹嘘自己提升了开发生产率。工程师可能确实写了更多代码，但他们要审核的代码比写的还多。现在的科技圈氛围就是加班文化加AI狂热。

Besides this being a wasteful measure, Arora gave an example of why it’s bad for a software org: “We had a software engineer producing 7X the code than anybody on her team. A superstar. Not only that, but also high-quality code. The check-ins and the reviews were awesome. But the other six people on the team were spending the majority of the time reviewing her code [rather] than writing the code.”

这种指标不仅毫无意义，Arora还举了个例子说明它对开发团队的危害：「我们团队有个软件工程师产出是其他成员的7倍，堪称明星员工。不光数量多，代码质量也很高，提交和初评结果都很好。但团队里其他6个人大部分时间都在审核她的代码，根本没时间写自己的代码。」

Code reviews require broad expertise in a codebase, especially if that review is to be effective and helpful. Reviewers have to look at the change in the context of the larger system, which requires holding and understanding the context of the larger system. That requirement to pass judgement on a code commit can cause a lot of stress on a reviewer. “You're essentially asked to contribute your expertise,” said a senior engineer now at Intuit. “So there is an element of, ‘If I mess up this review, I was the gatekeeper of this code. And if I mess it up, that might be my fault.’ So there's a lot of pressure there.”

代码审核需要审核者对整个代码库有全面的了解，尤其是要做出有效、有价值的审核的时候。审核者必须结合整个系统的上下文来看代码变更，这就要求他们掌握、理解整个系统的上下文。要对代码提交做出判断，会给审核者带来很大压力。「本质上是公司要求你输出专业判断，」现在在Intuit工作的一位资深工程师说，「所以你会有这种想法：‘如果我审核出了问题，我就是代码的守门人。要是出了错，可能就是我的责任。’所以压力特别大。」

Every builder is a decider

每个开发者现在都是决策者

Smartsheet’s research found that 80% of AI-generated content is edited before it’s finalized. Those edits come from getting an understanding of the context of that code (or other content). For AI-generated code, no one wrote the original code, so the context you need to gather is greater. You can look at prompts, specs, and whatever other context your agent uses, but that’s a lot of work to produce a judgement. If we’re shifting the majority of software work from coding to making decisions, everyone is going to feel the strain of decision fatigue.

Smartsheet的研究发现，80%的AI生成内容在最终定稿前都需要人工修改。这些修改建立在理解代码（或其他内容）上下文的基础上。对AI生成的代码来说，没有人写过原始代码，所以你需要收集的上下文信息更多。你可以去看提示词、需求文档、AI用到的其他上下文资料，但要做出判断需要做大量工作。如果我们把软件开发的大部分工作从写代码转移到做决策，每个人都会感受到决策疲劳的压力。

{{image_3}}

Before AI-enabled software engineering workflows, developers spent a significant percentage of their day (the exact figure varies wildly depending on who you asked) doing things other than writing code. Now that AI writes the code, there’s more time for the other work of an engineer. In fact, you may have seen the rise of the builder, which Arora defines as “anybody who understands a customer issue or a problem, has an idea who can prototype and build software quickly to test or ship.” The core skills of builders are understanding context and making judgement calls.

在AI进入软件开发流程之前，开发者每天有很大比例的时间（具体数字问不同的人差异很大）不是在写代码。现在AI负责写代码了，工程师就有更多时间做其他工作。实际上你可能已经注意到「Builder」这个角色的兴起，Arora把它定义为「任何懂客户问题、有想法、能快速搭建软件原型测试或上线的人」。Builder的核心技能就是理解上下文和做出判断。

Smartsheet and other industry studies have found that this shift doesn’t make developers’ lives easier; it makes them more intense. Multiple AI agents run in the background while the developer reviews code, attends meetings, and writes up documentation. They feel more productive, but aren’t always. “The hours haven't changed, but the density of work has, right?” said Arora. “The amount of decisions we're making in a day, how much information we are gathering and trying to make a decision out of it has changed.”

Smartsheet和其他行业研究发现，这种转变并没有让开发者的工作变轻松，反而让工作强度更高了。开发者审核代码、开会、写文档的时候，多个AI工具在后台运行。他们觉得自己生产率更高了，但实际并不一定。「工作时长没变，但工作密度变了，不是吗？」Arora说，「我们每天要做的决策数量，要收集、分析用来做决策的信息量，都完全不一样了。」

Conclusion

总结

Easy-to-generate code has meant harder-to-review pull requests. Those PRs need lots of context and lots of judgement, and developers are having to make decisions more often. That’s intense, and it’s leading to decision fatigue and burnout.

生成代码变简单，意味着合并请求的审核难度变高了。这些PR需要大量上下文信息和大量判断，开发者要做决策的频率越来越高。这种高强度的工作正在导致决策疲劳和职业倦怠。

AI problems might require AI solutions as the models, harnesses, and techniques improve. Would you trust an AI agent to build software end-to-end off of a single spec? Would you be okay giving up reviews on individual commits in exchange for a review on the final outcome? If software engineers and builders are going to work effectively in an AI-enabled world without throwing down their keyboards to become artisanal furniture builders, we may have to become comfortable with answering yes to both questions.

随着模型、工具链和技术的进步，AI带来的问题可能也需要AI来解决。你会信任AI工具根据一份需求文档端到端开发出完整软件吗？你愿意放弃审核每一次提交，换成只审核最终产出吗？如果软件工程师和开发者想要在AI时代高效工作，而不是扔了键盘去当手工家具匠人，我们可能迟早要对这两个问题给出肯定的答案。

原文来源：https://stackoverflow.blog/2026/05/21/coding-agents-are-giving-everyone-decision-fatigue/