正文 · AI 翻译
IT Home reported on October 28 that, on the 27th local time, Amazon updated its working guidelines for software engineers in the AI era, emphasizing that evaluations should consider how much business value engineers create, rather than how much AI they use.
Amazon Vice President and Distinguished Engineer Turk Tekowitz pointed out that, compared with earlier technological shifts such as cloud computing, AI is affecting software development faster and more profoundly. However, the key measure of engineers’ performance remains the results they deliver for customers and the business, rather than how often or how much they use AI. “We all have to adapt to change and spend time learning, or we risk being left behind.”
AI coding tools and agents are changing the software industry, and programmers’ work is changing with them. Amazon calls this phase the “Renaissance developer era” and believes that combining human judgment with machine intelligence can enable faster, more efficient development of new products and services.
The software industry is undergoing a transformation, and the professional role of software engineers is changing as well. In recent years, Amazon has placed increasing emphasis on engineers’ and technical managers’ use of AI and has tried to incorporate it into performance evaluations. The company has counted how often employees use AI, the number of tokens they consume, and the lines of code they write. It has also used some of these data to create leaderboards comparing different teams.
However, the memo warns that using these data to assess performance is different from asking employees to raise the numbers as much as possible. An excessive focus on metrics could backfire: employees might, for example, overuse AI to increase token consumption without producing better software or improving the customer experience. Tekowitz said, “If these metrics become targets rather than references for evaluating work, they could lead to negative outcomes.”
According to Business Insider, Amazon had previously been trying different approaches to both measure engineering teams’ AI use and encourage employees to use AI more.
Earlier this year, Amazon began tracking engineers’ use of internal AI tools, and some teams introduced leaderboards based on figures such as token consumption. The move drew complaints from some employees, who worried that unnecessary activity to meet AI usage requirements would be more likely to win recognition than genuine improvements in work efficiency.
The new guidelines are Amazon’s clearest public acknowledgment yet that it recognizes this risk. Amazon believes that detailed rules on how engineers should use AI could quickly become outdated as the technology advances, so the new guidelines focus on lasting results for customers and the business.
IT Home has learned that Amazon has not reduced its emphasis on adopting AI as a result. The company stresses that AI should be “at the core of software engineering practice” and requires engineers to use the technology in their daily work.
The memo also requires managers to use AI tools themselves on real projects, understand which problems AI can solve and where it still falls short, and explain the reasoning behind their judgments to their teams.
Tekowitz pointed out that Amazon has found some correlation between data on token consumption, AI tool use, and code reviews and the work outcomes the company values, but the numbers cannot be considered in isolation. Consuming more tokens can mean higher costs, while pursuing more lines of code for their own sake can make software unnecessarily complex.
For teams whose AI adoption varies substantially, managers need to examine 2 groups in particular: teams in the bottom 10%, to understand why they have not been able to use AI effectively, and teams in the top 10%, to determine whether extensive AI use has actually sped up delivery.
Amazon has found that the “frontier teams” using AI most actively have achieved better results in efficiency and delivery. Conversely, teams and employees that have not adopted AI or have failed to improve their output metrics are gradually falling behind others.
However, the new guidelines do not mean that every engineer must prove AI can improve their own efficiency. In roles that cannot currently benefit from existing AI tools, employees can document the specific reasons and discuss them with colleagues and managers to find solutions together. The memo says, “Engineers who can explain why AI cannot help for now are demonstrating exactly the judgment we want to see.”