Amazon releases engineer guidance document, opposing the use of Token consumption as a performance metric

In the past, the number of lines of code a programmer wrote may have been one of the metrics for measuring workload. Today, as AI programming tools become increasingly widespread, some companies have started fixating on a new number: Token consumption.

How many times is AI called each day? How many Tokens are consumed? How much code is generated? These figures seem to have become the new standard for judging whether programmers are actively embracing AI—but now Amazon has stepped forward to pour cold water on this trend: stop single-mindedly pursuing Token consumption; using AI a lot doesn't mean the work is done well.

Recently, Amazon publicly released a guidance document for software engineers, clearly stating that the company wants developers to focus more on the actual value AI creates, rather than blindly consuming Tokens in order to inflate usage figures.

Amazon reminds programmers: don't treat Token consumption as an achievement

This adjustment comes from Amazon's engineering management. In the official document, Terk Terkowitz, an Amazon Vice President and Distinguished Engineer, revealed that Distinguished Engineers, Senior Principal Engineers and other technical leaders from various business units across the company recently jointly revised the job guidance for Software Development Engineers (SDEs).

The core question under discussion was: AI is completely transforming software development, so should the standard for a good programmer change accordingly? Amazon believes the answer is yes, but developers cannot simply be defined by AI usage. Terk Terkowitz said:

He also proposed a new concept: the "Renaissance Developer," that is, an engineer who can combine AI tools, human judgment and domain expertise to create entirely new products.

One point needs to be made clear: Amazon is not opposed to employees using AI. On the contrary, the company believes AI will be at the core of future software engineering, and it also requires engineers who have not yet used AI to learn it as soon as possible and integrate it into their daily work—what really worries Amazon is another trend: treating the use of AI itself as a work achievement.

In the past, some Amazon teams would observe AI's impact on development efficiency through metrics such as Token consumption, AI tool adoption rate, number of code changes, and number of code reviews. These figures do have some reference value. Amazon found that "Frontier Teams" that actively embrace AI can usually achieve higher productivity and better delivery performance.

But the problem is that correlation does not equal causation, and these numbers certainly cannot be turned directly into performance targets. Amazon warned in the document:

"If these metrics are treated as targets rather than measurement tools, it can lead to negative outcomes, including higher costs and increased complexity."

For example, frequently calling AI in order to increase Token consumption will raise model usage costs; if the sole pursuit is generating more code, it may also make originally simple systems more complex. This runs counter to Amazon's long-emphasized principles of "Frugality" and "Invent and Simplify."

An Amazon spokesperson also confirmed to the media that one purpose of releasing this document is to discourage engineers from blindly pursuing Token consumption, the so-called "Tokenmaxxing."

In fact, Amazon has encountered similar problems before: according to Fortune magazine, an informal AI usage leaderboard called Kirorank once appeared inside Amazon, and some employees, in order to improve their rankings, would have AI Agents perform meaningless tasks to increase Token consumption, and Amazon shut down this leaderboard in May of this year.

However, regarding this adjustment, Amazon also emphasized: this does not mean engineers can stop learning AI. The official document clearly requires: employees who have not yet used AI should start learning as soon as possible, and employees who already use AI need to continue exploring better ways of working.

For company managers, Amazon also put forward a specific requirement: focus on observing teams whose AI adoption rate is around 10% and 90%.

The document also contains this sentence: "If your current work cannot benefit from existing AI tools, that is a discovery, not a failure." Amazon believes engineers should record the specific reasons why AI cannot help, share them with colleagues and managers, and not force themselves to use AI just for the sake of using it.

For technical leaders, Amazon's requirements are even more specific: they cannot just listen to others introduce AI; they must personally have used these tools in actual production work, be able to cite cases where AI played a role and where it did not, and explain the reasons to their teams.

In addition, Amazon also found a problem: as the company scales up, some engineers and managers gradually drift away from customers and the actual business; while teams that better understand the product, customer needs and business goals tend to be better able to leverage the advantages of AI and agent development.

Therefore, Amazon hopes engineers will not only know how to write code and call AI, but also clearly understand what problem they are solving: "After all, AI can help engineers write code and complete tasks, but to judge which problems are worth solving and which features truly meet customer needs, engineers still need a deep understanding of the product and business."

Microsoft also calls a halt to "burning Tokens": some people spend thousands of dollars a month

People who regularly follow this circle should know that Amazon is not the first tech giant to question Tokenmaxxing.

In August this year, Microsoft Executive Vice President Jay Parikh sent an email to internal development teams, asking engineers to use AI more rationally: "Tokenmaxxing is not our goal... I want everyone to focus on creating outcomes that truly drive customer and business development, and make every Token create greater value."

To this end, Microsoft immediately adjusted its internal AI usage guidelines, setting the cheaper GPT-5.6 as the internal default model, and requiring departments to manage Token spending more strictly—according to foreign media reports, at that time some Microsoft engineers could spend hundreds or even thousands of dollars a month on Token consumption alone.

Earlier, in June this year, AWS CEO Matt Garman also said that enterprises can choose models with appropriate costs based on specific tasks, rather than calling the most expensive AI for every task. For some daily development tasks, open-weight models that are "good enough" in capability and lower in price may be more cost-effective than top-tier closed-source models.

From Microsoft to Amazon, two tech giants successively putting the brakes on the trend of "frantically burning Tokens" is somewhat thought-provoking.

After all, not long ago they were still encouraging employees to use AI as much as possible, afraid that someone would fail to keep up with this wave of technology. Now, as Token bills become more and more expensive and AI-generated code becomes more and more abundant, management has begun to ask again: how much actual output did the money spent really buy?

For programmers, this may also be a somewhat contradictory situation: using AI less may be seen as failing to keep up with the times; using AI more may be questioned as wasting Tokens. So, is your company also tracking employees' Token consumption? Do you think this can truly reflect a programmer's work efficiency?

Reference links:

This article comes from the WeChat public account "CSDN", author: Zheng Liyuan, published by 36Kr with authorization.