亚马逊致力于改善构建者体验，并使用数据驱动的方法来提高员工生产力和满意度（体验）
早在 2022年，Business Insider 报道了一篇关于亚马逊 CEO Andy 创立了一个名为亚马逊软件构建者体验（ASBX）的新部门，以解决开发者提出的”基础性痛点”，Builder 或开发人员是很多科技大厂的核心资产和战斗力，以往我们一直在讲 DevOps 研发运维一体，亚马逊也是 DevOps 文化、机制和工具落地成功的典范公司之一，那为什么 2022年又成立这样的一个 ASBX 横向部门？亚马逊的构建者体验有哪些挑战？GenAI 又为开发者体验带来哪些新的机遇？
2022 内部技术生态和现状
https://www.aboutamazon.com/news/workplace/new-amazon-employee-experience-survey
2022年，ASBX 刚成立就接手了内部年度的技术调查和结果分析，并披露到 About Amazon 官网，所谓年度技术调查问卷，会发给所有的亚马逊技术员工，以了解和收集他们关于日常很多方面的各种反馈，看看 2022年的结果总结：
-
2022年技术调查问卷亮点包括：
- 员工满意度 （86%）和推荐亚马逊作为工作场所（81%）的得分很高
- 91% 的人感到被鼓励分享想法
- 74% 的人对在亚马逊实现职业目标充满信心
- 85% 的人报告经理在做决定时寻求多样化的意见
- 86% 的人说团队正在为客户做正确的事情，75% 反馈他们真正在为客户做创新
-
需要改进的领域：
- 5% 的受访者考虑在未来六个月内离开亚马逊，其中一半人将基本工资作为主要考虑因素
- 22% 的开发人员报告 bug 修复经常中断他们的工作
- 34% 的工程师每周花费 4-8 小时在非差异化工作上
- 0-20% 的工程师时间花在非产品构建任务上
- 30% 的工程师时间花在重复性任务上
- 内部工具与开源工具的兼容性不足，以及观察性工具的不足
-
ASBX 团队（成立于2022年2月）正在进行的改进包括：
- 自动解决 20% 的 Blocked Software Issues，即开发团队必须等待那些紧急重要的部署完成才能执行自身的部署
- 推出跨团队的管道效率仪表板 (Pipeline Efficiency Dashboard)
- 改进 SIM Ticketing 的搜索功能
- 集中处理 180万台机器的 Amazon Linux 2 的升级
- 减少大型集群 75% 的 Apollo 部署时间
ASBX 团队使命是什么？
https://www.businessinsider.com/amazon-builder-experience-team-uses-these-6-guiding-principles-2022-9
根据 Business Insider 的报道，“Amazon Software Builder Experience”(ASBX) 团队，成立初期有 400多人，目的是解决工程师们的不满，并培养更好的“构建者/开发者”文化，内部很多开发人员抱怨工作变得越来越重复和平凡，阻碍了开发人员从事更具创造性的活动；使亚马逊成为”地球上软件构建者最佳雇主”。
为了实现这样的目标，ASBX 团队制定了 6项指导原则，作为他们做出关键决策的核心价值观（信条）：
- 提供一致、可互操作和可扩展的工具
- 消除非差异化工作，实现自动化
- 确保工具在最糟糕的时期也能使用
- 通过指标、可行洞察和知识共享不断改进软件构建者体验
- 提供行业领先的技术和顶级专家资源，促进学习和成长
- 将亚马逊的价值观编码到技术基础中，培养包容性文化
这些原则旨在改善亚马逊工程师的工作体验，提高效率，并促进创新。
已知的工作内容包括但不仅限于：
- 开发和实施解决方案，如 Amazon Q，以提高开发人员效率
- 利用检索增强生成（RAG）技术结合亚马逊知识库，为开发人员提供快速、准确的答案
- 通过 S3 连接器将内部知识库导入 Amazon Q Business
- 对文档进行预处理和元数据丰富，以提高检索效率
2023年，有哪些提升？
通过 ASBX 披露的2023年技术调查的反馈，我们可以总结这一年的努力之后的结果：
- 建人员反映平均花费在非关键任务上的时间减少了 15%
- 由于未通过测试而导致的平均阻塞时间减少了 10%
- Pipeline 操作员干预次数 (一种衡量您需要手动解决部署 Pipeline 阻塞的指标) 下降了 30%以上
- 公司范围内的 sev2 工单减少超过 20%
ASBX团队引入了新的基准工具，帮助管理者更有效地分析数据。今年，他们进一步改进了这些工具，使其对所有人可用，并引入了新的基准评分系统，以便进行更相关的团队比较。
生成式 AI 浪潮汹涌而来
GenAI 辅助开发目前可以确定的一个拥有广泛前景的场景，包括 Microsoft Copilot，Amazon Q Developer，Gemini Code Assist 几个大厂和很多初创公司比如 Cursor AI，都在聚焦如何改善开发者体验；以下表格对比了 Q Developer 和 CoPilot 的差异，从差异也可以看到不同产品的聚焦方向；
| 领域/用例 |
Amazon Q Developer |
GitHub CoPilot |
| 定价 |
免费层次无时间限制 Q Professional 每用户每月 19 美元 |
GitHub CoPilot Business (每用户每月 19 美元) 可能缺乏功能开发和安全扫描 |
| 端到端功能开发和 SDLC |
内置,为所有软件开发角色提供整个 SDLC 的价值，不仅限于编码人员，根据高级描述生成可投入生产的代码 |
不适用 (私有技术预览) |
| 代码转换 |
是的 (免费层次和专业层次)；使用 Q Agent for Code Transformation (Java 可用，.Net 已宣布)。 |
不适用 |
| 安全扫描 |
✅ 原生包含 Java、JavaScript、Python 等 ✅ 可定制安全扫描 |
❌ 不包含 - 需要 GitHub 高级安全性 |
| AWS 服务知识和集成 |
✅ 与 AWS 平台深度集成,汲取亚马逊 17 年 AWS 最佳实践 - 建立在 Bedrock 之上 ✅ 指导选择正确的 AWS 服务、进行最佳配置以及排除服务相关问题 |
❌ 未原生包含 AWS 特定专业知识 |
| 对话能力 |
✅ 自然语言交互、回答架构问题、起草支持案例、支持解释 |
❌ 未原生包含,可能需要 CoPilot for Azure |
https://dev.to/aws-builders/how-amazon-q-stands-out-a-comparison-with-microsoft-copilot-and-google-gemini-1bj
Q Developer 有个独特的代码转换能力，可以帮助团队利用 LLM 大模型自动化升级 Java 应用和 .Net （.NET Framework 到跨平台 .NET 的升级，即将推出），ASBX 也是内部利用 Q Developer 构建自动化工具帮助团队升级到 Java 17，传统方法需要 50工程师人天，利用 Q Developer 升级一个 Java 项目只需要几个小时，在6个月时间内，ASBX 团队的工具帮助升级了超过 50% 的生产 Java 系统，亚马逊开发人员在直接交付了 79%自动生成的代码审查结果（Code Review），无需任何额外更改，该升级帮助公司提升效能相当于节约了 $2.6亿美金的成本，详情见 Andy 的 发布的信息：
https://www.linkedin.com/posts/andy-jassy-8b1615_one-of-the-most-tedious-but-critical-tasks-activity-7232374162185461760-AdSz/?utm_source%3Dshare%26utm_medium%3Dmember_ios
除了亚马逊自己，另外一个可查的代码转换公开案例是拉丁美洲的技术公司 Novacomp，”使用 Amazon Q Developer，Novacomp仅用 50分钟就升级了一个包含 10,000 多行 Java 代码的项目，而预计需要3周时间。该公司还简化了开发人员的日常任务,平均减少了 60% 的技术债务,并帮助客户显著改善了安全状况并节省了相关成本。”
你相信大语言模型可以大规模融入 SDLC 每个环节吗？我想大概还是需要更多先行者给大家更多信心和展示，不过个人而言，我觉得未来已来，我们要学习新的工具，习惯大语言模型在 SDLC 的每个环节的辅助、提效和体验改善。
想了解更多大语言模型在代码开发中的应用，可以参考2024年为 KDD 所做的一个调研报告 《Reasoning and planning with large language models in code development (survey for KDD 2024 tutorial)》
参考资料：
- https://careers.wct-fct.com/companies/amazon-3-60ad394d-c673-4474-9694-344b0cae748f/jobs/41486418-software-development-engineer-amazon-software-builder-experience-asbx
- https://gradle.com/blog/advice-for-andy-jassy-addressing-amazons-mammoth-developer-experience-challenge/
- https://www.aboutamazon.com/news/workplace/amazons-annual-tech-survey-results-now-available
Amazon is committed to improving the builder experience, using data-driven methods to enhance employee productivity and satisfaction.
Back in 2022, Business Insider reported that Amazon CEO Andy created a new department called Amazon Software Builder Experience (ASBX) to address “fundamental pain points” raised by developers. Builders or developers are core assets and fighting force for many tech giants. While we’ve always talked about DevOps integrating development and operations, Amazon has been one of the exemplary companies successfully implementing DevOps culture, mechanisms, and tools. So why establish this horizontal ASBX department in 2022? What challenges does Amazon face in builder experience? And what new opportunities does GenAI bring to developer experience?
2022 Internal Technology Ecosystem and Status
https://www.aboutamazon.com/news/workplace/new-amazon-employee-experience-survey
In 2022, shortly after its establishment, ASBX took over the internal annual technology survey and results analysis, which was disclosed on the About Amazon website. This annual technology survey is sent to all Amazon technical employees to understand and collect feedback on various aspects of their daily work. Let’s look at the 2022 results summary:
-
Highlights from the 2022 technology survey include:
- High scores for employee satisfaction (86%) and recommending Amazon as a workplace (81%)
- 91% felt encouraged to share ideas
- 74% were confident about achieving their career goals at Amazon
- 85% reported that managers seek diverse opinions when making decisions
- 86% said teams are doing the right things for customers, with 75% reporting they’re truly innovating for customers
-
Areas needing improvement:
- 5% of respondents considered leaving Amazon within the next six months, with half citing base salary as the primary factor
- 22% of developers reported that bug fixes frequently interrupt their work
- 34% of engineers spend 4-8 hours weekly on non-differentiating work
- 0-20% of engineers’ time is spent on non-product building tasks
- 30% of engineers’ time is spent on repetitive tasks
- Insufficient compatibility between internal tools and open-source tools, and inadequate observability tools
-
Improvements being made by the ASBX team (established in February 2022) include:
- Automatically resolving 20% of Blocked Software Issues, where development teams must wait for urgent important deployments to complete before executing their own deployments
- Launching cross-team Pipeline Efficiency Dashboard
- Improving SIM Ticketing search functionality
- Centrally managing the upgrade of Amazon Linux 2 on 1.8 million machines
- Reducing Apollo deployment time for large clusters by 75%
What is the ASBX Team’s Mission?
https://www.businessinsider.com/amazon-builder-experience-team-uses-these-6-guiding-principles-2022-9
According to Business Insider, the “Amazon Software Builder Experience” (ASBX) team started with over 400 people, aiming to address engineers’ dissatisfaction and foster a better “builder/developer” culture. Many internal developers complained that work had become increasingly repetitive and mundane, hindering them from engaging in more creative activities. The goal is to make Amazon “the best employer for software builders on Earth.”
To achieve this goal, the ASBX team established 6 guiding principles as core values (tenets) for making key decisions:
- Provide consistent, interoperable, and scalable tools
- Eliminate non-differentiating work through automation
- Ensure tools work even during the worst times
- Continuously improve software builder experience through metrics, actionable insights, and knowledge sharing
- Provide industry-leading technology and top expert resources to promote learning and growth
- Encode Amazon’s values into the technical foundation, fostering an inclusive culture
These principles aim to improve Amazon engineers’ work experience, increase efficiency, and promote innovation.
Known work includes but is not limited to:
- Developing and implementing solutions like Amazon Q to improve developer efficiency
- Utilizing Retrieval-Augmented Generation (RAG) technology combined with Amazon’s knowledge base to provide developers with quick, accurate answers
- Importing internal knowledge bases into Amazon Q Business through S3 connectors
- Preprocessing documents and enriching metadata to improve retrieval efficiency
What Improvements Were Made in 2023?
Through feedback from the 2023 technology survey disclosed by ASBX, we can summarize the results after a year of effort:
- Builders reported a 15% average reduction in time spent on non-critical tasks
- 10% reduction in average blocking time due to failed tests
- Over 30% decrease in Pipeline operator interventions (a metric measuring how often you need to manually resolve deployment Pipeline blockages)
- Over 20% reduction in company-wide sev2 tickets
The ASBX team introduced new benchmarking tools to help managers analyze data more effectively. This year, they further improved these tools, making them available to everyone, and introduced a new benchmark scoring system for more relevant team comparisons.
The Generative AI Wave is Surging
GenAI-assisted development is currently a scenario with broad prospects, including Microsoft Copilot, Amazon Q Developer, Gemini Code Assist from major tech companies, and many startups like Cursor AI, all focusing on improving developer experience. The following table compares the differences between Q Developer and CoPilot, highlighting the different focus areas of these products:
| Area/Use Case |
Amazon Q Developer |
GitHub CoPilot |
| Pricing |
Free tier with no time limit; Q Professional at $19 per user per month |
GitHub CoPilot Business ($19 per user per month) may lack feature development and security scanning |
| End-to-end feature development and SDLC |
Built-in, providing value across the entire SDLC for all software development roles, not just coders; generates production-ready code from high-level descriptions |
Not applicable (private technology preview) |
| Code transformation |
Yes (free and professional tiers); using Q Agent for Code Transformation (Java available, .Net announced) |
Not applicable |
| Security scanning |
✅ Natively includes Java, JavaScript, Python, etc. ✅ Customizable security scanning |
❌ Not included - requires GitHub Advanced Security |
| AWS service knowledge and integration |
✅ Deeply integrated with AWS platform, drawing on Amazon’s 17 years of AWS best practices - built on Bedrock ✅ Guides selection of the right AWS services, optimal configuration, and troubleshooting service-related issues |
❌ No native AWS-specific expertise |
| Conversational capability |
✅ Natural language interaction, answering architectural questions, drafting support cases, supporting explanations |
❌ Not natively included, may require CoPilot for Azure |
https://dev.to/aws-builders/how-amazon-q-stands-out-a-comparison-with-microsoft-copilot-and-google-gemini-1bj
Q Developer has a unique code transformation capability that helps teams leverage LLM to automate upgrading Java applications and .Net (.NET Framework to cross-platform .NET, coming soon). ASBX also uses Q Developer internally to build automation tools helping teams upgrade to Java 17. Traditional methods would require 50 engineer-days, while upgrading a Java project with Q Developer takes only a few hours. Within 6 months, ASBX team’s tools helped upgrade over 50% of production Java systems. Amazon developers directly delivered 79% of automatically generated code review results without any additional changes. This upgrade helped the company improve efficiency equivalent to saving $260 million in costs, as detailed in Andy’s published information:
https://www.linkedin.com/posts/andy-jassy-8b1615_one-of-the-most-tedious-but-critical-tasks-activity-7232374162185461760-AdSz/?utm_source%3Dshare%26utm_medium%3Dmember_ios
Besides Amazon itself, another publicly known code transformation case is Latin American technology company Novacomp: “Using Amazon Q Developer, Novacomp upgraded a project with over 10,000 lines of Java code in just 50 minutes, compared to an estimated 3 weeks. The company also simplified developers’ daily tasks, reduced technical debt by an average of 60%, and helped customers significantly improve their security posture and save related costs.”
Do you believe large language models can be integrated into every aspect of the SDLC on a large scale? I think we probably still need more pioneers to give everyone more confidence and demonstrations. Personally, though, I believe the future is already here. We need to learn new tools and get accustomed to large language models assisting, improving efficiency, and enhancing experience at every stage of the SDLC.
To learn more about the application of large language models in code development, you can refer to a research report prepared for KDD 2024: “Reasoning and planning with large language models in code development (survey for KDD 2024 tutorial)”
References:
- https://careers.wct-fct.com/companies/amazon-3-60ad394d-c673-4474-9694-344b0cae748f/jobs/41486418-software-development-engineer-amazon-software-builder-experience-asbx
- https://gradle.com/blog/advice-for-andy-jassy-addressing-amazons-mammoth-developer-experience-challenge/
- https://www.aboutamazon.com/news/workplace/amazons-annual-tech-survey-results-now-available
