软件只有在发布后才有价值；因此，加快发布节奏至关重要。在 Craft Conference 大会上，Titus Winters 发表了题为“软件工作流优化”的演讲。他指出，开发、部署和运维需要不同的思维方式、考核指标和投入。他提到了影响技术领域的四个方面：产品成功、计算能力、工程能力以及战略创新。他认为，最大的成本在于人力，而非计算资源，而人力成本往往未被充分衡量。他强调，人工智能促使工程师更深入地理解问题，而不是减少工程师的数量。
Winters 表示，无论是否使用人工智能开发软件，我们都必须仔细推敲软件工作流中每个阶段的目标、输入和输出。他提到了自己与 Leah Rivers 以及 Salim Virji 合著的论文“开发、部署、运维——理解软件开发成本与价值的整体模型”。该论文为推断软件流程的影响和效率提供了指导，无论该流程是否采用了人工智能。
Winters 认为，只有将软件交付给客户，它才有价值。他补充道，当生产环境中出现故障时，企业将面临实际的利润损失风险。能够通过快速、可靠地发布经过充分测试的更新来解决问题，本身就有价值。更快的发布节奏几乎总是更好，而且其好处可能比管理层所认为的还要大。
Winters 认为，开发是一个创造性的过程；这正是我们希望人类投入时间去做的事情。部署是一个工厂化流程，而运维则是一个自动化流程：
这三个流程不仅仅是不同的活动，它们需要不同的思维方式、不同的指标以及不同的投入。
Winters 提到了技术领域的四个根本性影响：
- 产品成功：收入、采用率、客户满意度、净推荐值（NPS）等
- 计算能力：云成本、存储、网络等
- 工程能力：生产力，即人力效率
- 战略创新：让新兴的好事物变得廉价/可行，让现有的坏事物变得罕见/不可行
Winters 提到，我们必须综合考虑这四个方面，并平衡它们的影响：
减少测试次数或许能节省计算成本，但由此引发的质量问题可能会导致实际的利润损失，或者在努力使产品达到稳定状态时造成工程师倦怠。
Winters 表示，最大的开销很少是计算成本，而是人力成本。我们无法直接衡量这些人员产出的成果，也无法准确地计算其产出的实际成本。他补充道，我们一直在衡量除最大开支项目以外的一切，这似乎并不严谨。
软件工程不仅仅是编写代码；它还涉及人和时间，而“人”这一部分是无法量化的。Winters 提到，随着编写代码的任务向人工智能转移，我们或许能够在设计实验时利用人工智能作为开发过程中“人”这一部分的替代方案。
Winters 认为，我们希望工程师主要专注于解决问题。借助人工智能，开发人员仍然将参与其中，但不再处于完全不受约束的状态。他们将在一个已经将专家知识和组织策略代码化的系统中工作，不再需要将所有细节都铭记于心，而是可以专注于眼前的独特问题。他总结道，我们所期待的转变并非减少工程师的数量，而是将关注点从单纯的软件专业技能转向对实际待解决问题更深入的理解。
InfoQ 在演讲结束后对 Titus Winters 进行了专访。
InfoQ：管理工程产能面临哪些挑战？我们该如何应对这些挑战？
Titus Winters：与追踪 token 成本或云服务支出相比，追踪工程师的时间都花在了哪里要困难得多。即使对于那些确实做了追踪的公司来说，我们也需要对工程流程有足够的了解，才能判断某项任务是否属于浪费，而这需要一定的敏锐度。
我们原本在阐明“开发者体验”与“生产力”是两个重要且相互关联的议题方面取得了良好进展，但人工智能似乎让这一讨论失去了活力。这并非因为必要的投入发生了变化，而是因为相关讨论变得过于混乱。在进入 AI 时代之后，这些经验教训依然成立，只是进程被加速了：我们需要清晰的流程、干净的代码、优质的工具、快速的反馈信号、完善的测试等等。
InfoQ：我们如何利用人工智能来设计实验，评估开发过程中的人为因素？
Winters：如果我们认可“token 消耗”可以作为软件任务中人力投入的可接受替代指标，那么我们就可以着手设计实验，为软件的各个方面赋予定量的成本或收益。
我们需要一套针对软件智能体的稳健的评估方案：涵盖多种软件任务（新功能开发、Bug 修复、重构），涉及我们关注的各类平台（前端、后端、移动端等），支持我们关注的多种编程语言，并包含不同难度的任务。该评估套件中的每项软件任务都需要一套功能和质量验收标准，而且这些标准需对人工智能不可见。此外，我们还需要能够测量每次运行评估的 token 成本。
在这样的设置下，我们现在可以进行实证测试了。你想验证良好的单元测试覆盖率是否真的有益吗？准备两个智能体：一个被提示要遵循我们所有的单元测试实践，另一个则主动不编写单元测试。运行评估套件，可能需要反复执行。测量 token 消耗量。重要的实践应该能全面地降低 token 成本；而专门化的实践可能仅在评估空间的某些区域里适用。
同样的原理也适用于工具。语义索引能否帮助人类浏览大型代码库？更具体地说，能带来多大帮助？让实验智能体使用一个优质的索引器，而控制组智能体则保留其标准的开箱即用方法。虽然我们需要多消耗一些 token 来解释如何使用这些工具，但如果该工具在净 token 成本上能带来回报，我们就应该投资于那些工具。
原文链接：https://www.infoq.com/news/2026/10/optimize-software-workflow/
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Software is only valuable after publish; it is therefore essential to speed up publish. At Craft Conference, Titus Winters gave a speech entitled “Software workflow Optimization”. He noted that development, deployment and mobility required different ways of thinking, testing indicators and inputs. He referred to four areas that affected technology: product success, computing capacity, engineering capacity and strategic innovation. In his view, the greatest cost lay in human resources, not in the calculation of resources, which were often not adequately measured. He stressed that artificial intelligence enabled engineers to understand the problem more deeply than to reduce the number of engineers.
Winters indicates that whether or not the software is developed using artificial intelligence, we have to refine the objectives, input and output of each phase of the software workflow. He referred to his paper with Leah Rivers and Salim Virji, “Developing, deploying, managing — understanding the overall model of the cost and value of software development”. The paper provided guidance for extrapolating the impact and efficiency of the software process, whether or not the process uses artificial intelligence.
Winters believes that the software has value only if it is delivered to the client. He added that when there was a malfunction in the production environment, the enterprise would face a real loss of profits risk. It is valuable in itself to be able to resolve the problem through a fast and reliable, fully tested update of publish. The faster publish rhythm is almost always better and its benefits may be greater than management believes.
Winters believes that development is a creative process; that's exactly what we want human beings to invest in. Deployment is a plant-based process, while transportation is an automated process:
These three processes are not just different activities; they require different ways of thinking, different indicators and different inputs.
Winters mentions four fundamental impacts in the field of technology:
- Product success: income, adoption rate, client satisfaction, net recommended value (NPS), etc.
- Computation capacity: cloud cost, storage, network etc.
- Engineering capacity: productivity, i.e. human efficiency
- Strategic innovation: making new good things cheap/feasible, making existing bad things rare/unfeasible
Winters mentioned that we must consider these four aspects together and balance their impact:
A reduction in the number of tests may result in cost savings, but the resulting quality problems may result in actual loss of profits or in the fatigue of engineers in their efforts to stabilize the product.
Winters states that the largest cost is rarely cost, but human cost. We cannot measure directly the results of those outputs or accurately calculate the actual costs of their outputs. He added that we had been measuring everything except the largest item of expenditure, which did not seem to be strict.
Software engineering is not just a code; it also involves people and time, and the “person” component cannot be quantization. Winters mentioned that as the task of writing codes shifts to artificial intelligence, we may be able to use artificial intelligence as an alternative to the human component of the development process when designing experiments.
We want engineers to focus on solving problems. Using artificial intelligence, developers will remain engaged, but no longer remain completely unfettered. They would work in a system that already coded expert knowledge and organizational strategies, and would no longer need to keep all the details in mind, but could focus on the unique issues at hand. He concluded that the shift we were looking for was not to reduce the number of engineers, but to shift the focus from mere software expertise to a deeper understanding of what was actually going to be done.
After the presentation, InfoQ gave an interview with Titus Winters.
InfoQ: What are the challenges in managing engineering capacity? How should we address these challenges?
Titus Winters: Tracking engineer time is much more difficult than tracking token costs or cloud service expenditures. Even for those companies that do track, we need to have sufficient knowledge of the engineering process to judge whether a task is wasteful, which requires some sensitivity.
We had made good progress in clarifying that “developmenter experience” and “productivity” were two important and interrelated topics, but artificial intelligence seems to have made this discussion unviable. This is not because the necessary inputs have changed, but because the discussions have become too confusing. After entering the AI era, these lessons remain valid, although the process has been accelerated: we need clear processes, clean codes, quality tools, quick feedback signals, good testing, etc.
InfoQ: How do we use artificial intelligence to design experiments to assess human factors in development?
Winters: If we accept that “tokeen consumption” can be an acceptable proxy for human input in software assignments, then we can begin to design experiments that assign quantitative costs or benefits to all aspects of software.
We need a robust assessment of the software AI agent: covering a wide range of software tasks (new functionality development, bug repair, remodelling), involving a variety of platforms of our concern (front, back, mobile, etc.), supporting the variety of programming languages that we are concerned with, and including different difficult tasks. Each software task in the assessment package requires a set of functionality and quality acceptance criterias, and these standards need to be invisible to artificial intelligence. In addition, we need to be able to measure the cost of the token per operation assessment.
With this setup, we can now do an empirical test. Do you want to verify whether good unit testing coverage is really useful? Preparation of two AI agents: one was advised to follow all our module testing practices, while the other took the initiative not to prepare the module testing. Operating the evaluation package, which may require repeated implementation. Measure token consumption. Important practices should reduce token costs in a comprehensive manner; specialized practices could be applied only in certain areas of space assessment.
The same rationale applies to tools. Can semantic indexes help humans view large code libraries? More specifically, how much help can it bring? The experimental AI agent was allowed to use a high-quality indexer, while the control group AI agent retained its standard open-box method. While we need to consume more token to explain how to use these tools, if the tool pays off on net token costs, we should invest in them.
Original link: https://www.infoq.com/news/2026/10/optimize-software-workflow/
Auto-fetched and translated by AI Atlas. All rights belong to the original author.