[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-harvard-coding-agent-review-bottleneck":3,"topics-all":41,"news-related-da6cbb98-3b89-4631-8dbd-59411136fb30":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"da6cbb98-3b89-4631-8dbd-59411136fb30","哈佛用 700 家公司拆解 AI 编程智能体:30% 代码增量,49% 合并时延","哈佛 Chen\u002FStratton 工作论文基于 700 多家公司 3 亿条工程事件显示,AI 编程智能体让代码量上升 30%、PR 数增加 23%,但 Jira 完成率无显著变化,且 PR 合并时间膨胀 49%,审查瓶颈并未消失,只是被搬到下一站。","十月初,Solidot 转载了一条来自 Ars Technica 的报道,把 AI 编程工具一年来的\"生产力幻觉\"摊在了桌面上。一份由哈佛经济学家 Fiona Chen 与 James Stratton 合作的工作论文,把 Jellyfish 工程遥测和 GitHub 活动数据并在一起,横跨 700 多家软件开发公司、约 70 万员工、3 亿条工作事件,从 2021 年一直追到 2026 年 3 月,想要回答一个关键问题:AI 编程智能体进入企业之后,公司到底有没有交付更多软件?\n\n## 数字很直接:代码起来了,产出没起来\n\n论文给出的结论干净到近乎残酷。引入 AI 编程智能体(覆盖 Claude Code、Cursor、Devin 这一类工具)之后,公司的代码总行数平均增加 30%,commit 数量增加 20%,pull request 数量平均增加 23%。这本来应该是\"生产力大幅提升\"的剧本。\n\n但同一组数据里另一些数字并不配合:Jira 里的 issue 完成率与 epic 完成率,在统计意义上没有显著变化。也就是说,写代码变得更快,但交付的功能没有变多。\n\n瓶颈没有消失,只是被搬到了下一站。论文披露,引入 AI 智能体之后,从 PR 提交到合并的平均耗时上升了约 49%,PR 上的评论数量上升了约 35%,revert 与返工的次数几乎翻倍。到 2026 年 3 月,80% 的公司已经启用了 AI 代码评审工具,但这些 AI 评审工具贡献的评论占比只有 23%——绝大部分代码审查工作依然由人类完成,且这部分人类工作时间被显著拉长。\n\n## 三个值得展开的细节\n\n第一,**写代码便宜了,但审查昂贵了**。AI 智能体能用 5 到 30 倍于普通聊天的 token 来跑\"先修后测\"的循环,这部分成本在 Jupyter \u002F Jellyfish 这类遥测里被算成 PR 提交数,而评论、返工、rebase 这些下游成本却发生在工程师身上。换句话说,Gartner 在 8 月警告的\"agent token 用量 5–30 倍\"不是抽象命题,它直接把审查队列拉长了。\n\n第二,**Jira 维度比 commit 维度更接近真问题**。Jira 的 issue 与 epic 是和具体业务目标绑定的;commit 与 PR 数量则是流程产物。Chen 与 Stratton 的论点是,后者的上升在管理学意义上是一种 \"速度\",而前者才是\"结果\"。当团队把代码行、PR 数写进 OKR 时,这层错位会被加倍放大。\n\n第三,**AI 代码评审工具的实际渗透,远比营销叙事保守**。80% 公司上线了,但只有 23% 的评论是 AI 写的。这说明审查这一环节里 AI 还没有\"能用\",处于\"上线了但用不起来\"的尴尬状态。\n\n## 这件事真正刺痛的,是公司在量什么\n\n论文没有否定 AI 智能体的能力,它在意的,是组织拿什么指标在衡量它。当一家公司在仪表盘里写的是\"PR 数 \u002F commit 数 \u002F 代码行数\",AI 智能体会很漂亮地完成所有指标,但 Jira 里的真实进度会纹丝不动。反过来,把\"每个 issue 的成本\" \u002F \"从 PR 到 merge 的时间\" \u002F \"merge 后的回滚率\"当成主指标,公司才能看到 AI 到底买到了什么。\n\n这里其实有一个更深的信号:agent 改变了工作的瓶颈形状。以前瓶颈在\"一个人写多快\",现在瓶颈在\"评审队列多深\"。如果一个组织不去重新分配人力,不让人类做合并、强制推送、删除分支这些不可逆动作,反而默认让 AI 智能体接管审查,那 49% 的合并时间膨胀只会更糟。\n\n## 所以这是什么级别的研究\n\n从 Solidot 转载的措辞看,这是工作论文,尚未经过同行评审。它的方法学强度来自样本规模——700 多家公司、3 亿事件,这套数据比任何单家公司的对照实验都有说服力,即便它使用的是 Jellyfish 自己的可观测性数据,样本选择会有偏差。\n\n但即便带着这些保留意见,这份论文是当下\"AI 编程工具到底有没有用\"这个问题里,横跨面最广、争议最少的一次系统回答。它告诉我们:接下来一两年,真正能拉开差距的,不是用不用 AI 编程智能体,而是谁先把审查流程和度量体系补回来。\n\n参考来源:\n\n- [Ars Technica: AI coding agents generate more code, but not more software](https:\u002F\u002Farstechnica.com\u002Fai\u002F2026\u002F10\u002Fai-coding-agents-generate-more-code-but-not-more-software\u002F)\n- [Solidot 转载: AI 编程智能体生成了更多代码,但并没有生成更多软件](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85589)\n","https:\u002F\u002Farstechnica.com\u002Fai\u002F2026\u002F10\u002Fai-coding-agents-generate-more-code-but-not-more-software\u002F","2af9d198-9418-4f26-85e4-4a8f3eede35a",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":25,"name":26,"slug":26,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"d3e8bb67-1c70-40d6-a0fa-a488171a5f32","en","Harvard Study: AI Coding Agents +30% Code, +49% Merge Delay","Harvard paper on 700+ firms: AI coding agents add 30% LOC, 23% PRs. Jira completion flat. Merge time balloons 49% despite 80% AI review adoption.","Early October, Solidot relayed an Ars Technica story that laid bare what people inside engineering orgs had been suspecting: AI coding tools may be more productive at producing code than at producing software. A Harvard working paper by economists Fiona Chen and James Stratton stitched Jellyfish engineering telemetry to GitHub activity signals across more than 700 software companies, around 700,000 developers, and roughly 300 million work events from 2021 through March 2026, asking a single question: did companies ship more software after AI coding agents landed?\n\n## The numbers are blunt: code up, output unchanged\n\nOnce firms adopted AI coding agents — Claude Code, Cursor, Devin, and friends — total lines of code rose about 30 percent on average, commits 20 percent, pull requests 23 percent. By every traditional developer-productivity metric, that should be a win.\n\nBut the same dataset shows Jira issue resolution and epic completion did not move in any statistically significant way. Writing code got cheaper. The shipped feature did not.\n\nThe bottleneck did not disappear — it moved downstream. Time from PR submission to merge climbed by roughly 49 percent. Comments per pull request climbed roughly 35 percent. Reversion and follow-up fixes nearly doubled. By March 2026, 80 percent of firms had AI code-review tools live, yet only 23 percent of review comments came from an AI; the rest, and the longer hours, still landed on humans.\n\n## Three details worth dwelling on\n\nFirst, **writing got cheaper; reviewing got expensive.** AI coding agents use 5 to 30 times more tokens than a chatbot for an equivalent task. That cost shows up as commit volume, while the comment churn, the rebases, and the back-and-forth show up on engineers’ timesheets. Gartner’s August warning about agent inference costs rising fivefold inside two years is not abstract — it is exactly what a 49 percent merge-time blow-up looks like when nobody re-pools the workflow.\n\nSecond, **Jira is closer to truth than GitHub.** Issues and epics are tied to business outcomes. Commits and PRs are artefacts of process. The authors’ central claim is that the latter rises while the former does not, and the disconnect gets worse when managers reward the latter as \"velocity\".\n\nThird, **AI code review tooling is far less adopted than the marketing implies.** 80 percent of firms have it on; 23 percent of comments come from it. Review is the chokepoint where AI is still \"live but unusable\".\n\n## What actually hurts is what the company chooses to measure\n\nThe paper does not argue that AI coding agents are weak. It argues that organisations are measuring the wrong thing. As long as the dashboard reads PR count, commit count, and lines of code, every coding agent becomes a beautiful performer on the wrong stage. Move the metrics to cost-per-resolved-issue, PR-to-merge time, and post-merge rollback rate, and the same agents suddenly look very different in different shops.\n\nThere is a deeper signal here. Agents change the shape of the bottleneck. The bottleneck used to be how fast one person can type; it is now how deep the review queue is. An organisation that does not reallocate human attention to the irreversible actions — merging, force-pushing, deleting branches, editing repo settings — and instead lets agents absorb review as well, will see that 49 percent extend further, not contract.\n\n## How seriously to take this\n\nThe piece, as relayed by Solidot, is a working paper, not peer-reviewed. Its empirical strength comes from the sample size: 700 firms, 300 million events is a degree of breadth no single-company A\u002FB can match. Its weakness is selection bias: every company in the dataset runs Jellyfish tooling, which biases toward mid-market software firms that already buy engineering analytics.\n\nEven with those caveats, this is the most cross-sectional and least ideological answer to \"do AI coding tools actually work?\" that has surfaced this year. The takeaway is clean: the next twelve to twenty-four months of differentiation will not come from whether a team uses AI coding agents. It will come from who redesigns the review queue and the metrics first.\n\nReference:\n\n- Ars Technica: AI coding agents generate more code, but not more software — https:\u002F\u002Farstechnica.com\u002Fai\u002F2026\u002F10\u002Fai-coding-agents-generate-more-code-but-not-more-software\u002F\n- Solidot relay (Chinese) — https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85589\n","harvard-coding-agent-review-bottleneck","2026-10-14T11:30:00Z","2026-10-14T11:06:33.228099Z","2026-10-14T11:06:33.228109Z",true,"agent",314,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"859242dd-0035-41e6-8db9-4583989048b7","AgentSWE基准:让编码智能体造智能体,无人及格","agentswe-coding-agent-benchmark","2026-10-13T19:08:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"c368ad9f-9308-4a0b-8f5c-3ae4601b48b9","D2K-Bench: 专家设计把 LLM 写 GPU 核提速 33.9%","d2k-bench-llm-gpu-kernel-design-guidance","2026-10-07T03:00:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"85f7f1c2-d896-436b-a915-37faed8776eb","你改主意了,模型没改:被拒需求也会带偏大模型","intent-eval-rejected-change-confusion","2026-10-06T17:15:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"4c4e444f-9614-42fe-b8f4-743204f854fa","RL 后训练收「锐化税」:base 模型配轻 harness,pass@K 反超官方版","sharpening-tax-rl-post-training","2026-10-03T15:09:45+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00"]