我们现在已进入一个应当对彼此抱有 3 倍更高期望的时代。
过去六个月里,一个又一个数据点接踵而至:
- NVIDIA 报告称,3 万名开发者的代码提交量增长了 3 倍,而缺陷率保持平稳。¹
- Amplitude 的每周生产环境代码提交量增长了两倍,目前一个 AI 智能体已成为代码库的前三大贡献者之一。²
- Anthropic 在内部采用 Claude Code 后,测得每位工程师编写的代码量提升了 2.5 倍,且质量保持稳定。³
- Replit 在同一时期内团队规模翻倍,人均产出增长了两倍,而评审时间、回退次数和事故数量均保持平稳。⁴
上图将整个生态划分为三个不均衡的层级,每个层级由公司捕获模型能力的程度来界定。⁵
第一个层级是如今大多数公司的现状。分发一个 AI IDE,其他什么都不改,结果只能算是平平。
“工程领导者带着 2-3 倍生产力提升的预期投入 AI,但实际落地效果却更接近 30%。”
—— Augment Code⁶
Faros 对 22,000 名开发者的遥测数据证实了这一点:工程师完成史诗级任务的效率提升了 66%,但每位开发者的缺陷数却增加了 54%。⁷ Google 的随机对照试验给出的数字是 21%,与 GitHub 的 24% 接近。⁸⁹ 这就是默认结果。
接下来是前沿层级。这里的公司围绕模型构建了完整的“驾驭框架”,编排能够在 GitHub、Linear 和 Slack 之间共享上下文的智能体;并在需要判断时升级给工程师处理。
“每位员工都会得到一个管理智能体,它会在循环中派生出工作智能体。我们的内部智能体在安全测试和事件分诊方面的表现超过了七位数的 SaaS 工具,而成本仅为后者的十分之一。”
—— Amjad Masad,Replit,《自动驾驶公司》⁴
人工 PR 评审时间下降了 30%。复杂支持处理时间下降了 60%。总代码贡献量提升了 5.8 倍。3 倍这个数字就存在于这个层级。
第三个层级是软件工厂,在这里这个名称恰如其分。它们是机械化生产软件的 AI 机器。Cognition 的 Devin 能够端到端地重构单体代码库。Factory.ai 正在 NVIDIA、Adobe、Blackstone 和 EY 部署软件工厂。¹⁰
Nubank 使用 Devin 进行大规模重构,工程效率提升了 8 倍,成本降低了 20 倍。
——Contrary Research,2026 年 1 月 11
高盛正在与 12,000 名人类开发人员一起试点 Devin,并公开估计智能体 AI 的效能提升速度可能是以往工具的 3-4 倍。12
AI 带来的工程生产力提升已经到来。初步数据揭示了可预期的结果:大多数团队应从 20% 的生产力提升跃升至 3 倍的生产力提升,而且这些提升并非寻常水平。
-
Cursor,“NVIDIA 如何使用 Cursor”,2026 年 2 月。↩︎
-
Cursor,“Amplitude 与 Cursor 云智能体”,2026 年 4 月。↩︎
-
Boris Cherny,Claude Code 负责人,Big Technology 播客,2026 年 7 月。↩︎
-
Amjad Masad,“自动驾驶公司”,2026 年 7 月 16 日。↩︎ ↩︎
-
上述分布为示意性质,并非统计数据。每个数据点均来自已发表研究、随机对照试验或公司披露的倍数报告。该分布并非取自抽样总体,曲线是对报告结果模式进行右偏对数正态拟合的结果,而非对原始数据的拟合。请将其视为一种形态论证,而非估算工具。↩︎
-
Augment Code 在 X 平台上的发文,2026 年。↩︎
-
Faros,“2026 年 AI 工程报告”。↩︎
-
Google 内部随机对照试验,约 100 名工程师,2024 年。参考自 DORA 报告;综述见 Value Add VC。↩︎
-
GitHub、微软和埃森哲与一家大型金融科技公司合作开展的研究,约 450 名开发人员,2024 年。↩︎
-
Factory.ai,“Factory 2.0:从编码智能体到软件工厂”。↩︎
-
Contrary Research,“Cognition”,2026 年 1 月。↩︎
-
CNBC,“高盛试点其首个自主编码器,成为华尔街 AI 领域的重要里程碑”,2025 年 7 月。↩︎
We are now in an era where we should expect 3x more from each other.
Over the last six months, one data point has followed another :
- NVIDIA reported a 3x increase in committed code across 30,000 developers with bug rates flat.1
- Amplitude tripled weekly production commits, with an AI agent now a top-three contributor to the codebase.2
- Anthropic measured a 2.5x increase in code written per engineer since adopting Claude Code internally, quality stable.3
- Replit doubled its team & tripled per-engineer output over the same period, with review times, reversions, & incidents all flat.4
The chart above sorts the ecosystem into three unequal tranches, each defined by how much of the model’s power the company captures.5
The first tranche is what most companies experience today. Distribute an AI IDE, change nothing else, & the outcome is modest.
“Engineering leaders went into AI expecting 2-3x productivity gains but are landing closer to 30%.”
— Augment Code6
Faros’s telemetry across 22,000 developers confirms this: engineers completed epics 66% faster, but bugs per developer increased by 54%.7 The Google randomized controlled trial put the number at 21%, close to GitHub’s 24%.89 This is the default outcome.
The frontier tranche follows. Companies here have built harnesses around the model, orchestrating agents sharing context across GitHub, Linear, & Slack; escalating to engineers for their judgment.
“Every employee gets a manager agent that spawns worker agents in loops. Our internal agent outperformed a seven-figure SaaS tool in security testing and incident triage at one-tenth the cost.”
— Amjad Masad, Replit, “The Self-Driving Company”4
Human PR review time dropped 30%. Complex support handling time dropped 60%. Total code contribution rose 5.8x. This is where the 3x number lives.
The third tranche are the software factories, & here the name is an apt descriptor. They are AI machines that produce software mechanistically. Cognition’s Devin refactors monolithic codebases end-to-end. Factory.ai is deploying software factories at NVIDIA, Adobe, Blackstone, & EY.10
“Nubank achieved an 8x improvement in engineering efficiency & a 20x cost reduction using Devin for large-scale refactoring.”
— Contrary Research, January 202611
Goldman Sachs is piloting Devin alongside 12,000 human developers & publicly estimates agentic AI could deliver 3-4x the rate of prior tools.12
AI engineering productivity gains are here. The initial data shows what to expect: most teams should migrate from 20% productivity gains to a 3x productivity gain & they aren’t normal.
-
Boris Cherny, head of Claude Code, on the Big Technology podcast, July 2026. ↩︎
-
Amjad Masad, “The Self-Driving Company,” July 16, 2026. ↩︎ ↩︎
-
The distribution above is illustrative, not statistical. Each point is a reported multiplier from a published study, RCT, or company disclosure. It is not drawn from a sampled population, & the curve is a right-skewed log-normal fit to the pattern of reported outcomes, not to raw data. Treat it as a shape argument, not an estimator. ↩︎
-
Google internal randomized controlled trial, ~100 engineers, 2024. Referenced in DORA reports; roundup at Value Add VC. ↩︎
-
GitHub, Microsoft, and Accenture study with a large fintech, ~450 developers, 2024. ↩︎
-
Factory.ai, “Factory 2.0: From coding agents to software factories”. ↩︎
-
Contrary Research, “Cognition”, January 2026. ↩︎
-
CNBC, “Goldman Sachs is piloting its first autonomous coder in major AI milestone for Wall Street,” July 2025. ↩︎