Skip to content
audiotale.net
Sign in
NewsNews

Amazon memo to engineers: tokenmaxxing is at odds with frugality, AI non-use is 'a finding, not a failing'

An Amazon VP's engineering memo calls token counts and lines of code deeply flawed measures, yet says AI usage still correlates with outcomes.

audiotale.net team

· 6 min read

Written by AI from the sources linked below. How we use AI

On this page
  1. Key points
  2. What happened
  3. Background
  4. How the guidance works in practice
  5. What it means
  6. What we don't know yet
  7. What to watch next
  8. Sources

Amazon has made public an engineering memo that tells its software developers not to chase AI usage numbers for their own sake. In the memo, a vice president and distinguished engineer calls token counts and lines of code "deeply flawed" ways to measure productivity, warns that "tokenmaxxing" clashes with Amazon's frugality principle, and still says engineers who ignore AI risk being left behind. The message is mixed on purpose, and that is what makes it worth reading closely if your own employer is counting tokens.

Key points

  • Terk Terkowitz, a VP and distinguished engineer at Worldwide Amazon Stores, wrote the memo to Amazon's software development engineers (SDEs) and managers (SDMs); Amazon published it on its news site, and Fortune reports it went public on Tuesday 27 October 2026.
  • It says "neither the developer role nor their output is defined by the use of AI", and "If your work does not benefit from current AI tooling, that is a finding, not a failing."
  • It calls input metrics such as token usage and tool adoption, and throughput metrics such as code changes, flawed, and says "tokenmaxxing is at odds with frugality."
  • It still says input metrics correlate well with outcomes so far, and tells managers to look at the 10th and 90th percentile of AI use on their teams.
  • The memo gives no figures for token usage or AI adoption.

What happened

Amazon posted "An update from Amazon engineering leadership on the role of the developer in the age of AI" on its company news site. The post introduces a memo addressed to SDEs and SDMs that revises Amazon's SDE role guideline. According to The Register, the memo was first sent to engineers and managers in September 2026; Fortune dated its own report 28 October and says Amazon published the memo on the Tuesday before. The Amazon page itself carries no visible date, so the exact publication date rests on Fortune's account.

The memo makes several points. It says Amazon has historically kept its engineering guidelines neutral about specific technologies, that leaders debated whether AI is different from cloud, microservices and serverless, and that they concluded it changes the work faster and more fundamentally. It says it is still early and that everyone should learn AI and apply it to daily work now.

It then explains why it does not define AI best practices: doing so risks "freezing our engineering practices in amber", so the guideline focuses on what stays constant, namely impact on customers, the business and teams.

On measurement, it acknowledges that input metrics (token usage, tool adoption) and throughput metrics (code changes, reviews) are flawed, and warns that treating them as targets could raise costs and complexity. It names tokenmaxxing as at odds with frugality, and measuring lines of code as at odds with inventing and simplifying. Yet it also says these inputs have correlated well with outcomes so far and are expected to for some time, and that they may be used "with care and judgement" (Fortune's quote).

For managers it proposes comparing the 10th percentile of AI use on a team, to understand why those people are not adopting it, with the 90th percentile, to check whether delivery rises with heavy use. It asks leaders to use AI hands-on for production work, and asks engineers at all levels to stay close to customers and product direction. It also says Amazon has seen some individual contributors and managers drift from customers and the business, and wants to reverse that.

Background

"Tokenmaxxing" is the habit of burning through AI tokens to look productive, usually because usage is visible on a dashboard. Fortune reports that employees at Amazon ran AI agents on meaningless tasks to inflate their counts on an unofficial internal leaderboard, which it calls Kirorank, and that Amazon ended it in May over concerns about computing costs. The Register, which spells it KiroRank, describes it as built by an employee and shut down in May 2026 after people created junk tasks to climb it. The two accounts agree on the outcome and the month and differ on the stated reason, so we describe both.

The memo lands in a year when AI use is a headline workplace measure. Our own earlier coverage of GitHub Octoverse 2026 reported that agents start about 40% of pull requests, one reason managers pay attention to usage. Fortune also quotes Shopify CEO Toby Lütke as saying last month that some workers produce "slop grenades", AI-generated emails and code that create more review work for colleagues; we did not check that statement against a Shopify source.

How the guidance works in practice

Reading the memo as a set of rules, four things stand out.

First, outcomes beat activity. The role is defined by results for customers, not by how many tokens someone spent, and the memo says so for both input and throughput measures.

Second, non-use has to be explained, not hidden. The memo asks engineers who find AI does not help a job to document why, share it with peers, raise it with their manager, and look for solutions. Fortune summarises this as telling management when AI is not useful and why.

Third, managers are told to look at both ends of the distribution. The bottom end shows who is not adopting AI and why; the top end shows whether heavy users deliver more.

Fourth, the memo keeps adoption on the agenda. It says everyone should learn AI now, and it says teams that embrace AI most are "frontier teams".

What it means

For developers at companies that track AI usage, the memo is a useful reference. It is an explicit statement from a large employer that a high token count is not a goal and that "this does not help my work" is acceptable when you can say why. It does not remove the pressure: it still treats adoption as correlated with outcomes and tells managers to watch the 10th percentile.

Corey Quinn, writing in The Register, reads it more sceptically. His column argues the message is mixed because it accepts justified non-use while also warning people they will fall behind, which he says could let a manager justify either outcome. He also points out that AI vendors, including AWS, bill by token consumption. That is his opinion, not a reported fact, and the memo does not address it.

For companies copying the approach, the practical content is the percentile check: pair a usage number with a delivery number before judging anyone.

What we don't know yet

  • Whether Amazon has changed how it reviews or rates engineers in light of the memo. The memo is a guideline; we saw no reporting on performance review changes.
  • The memo gives no numbers for token use or adoption, so how widespread tokenmaxxing was at Amazon is unknown beyond the leaderboard reports.
  • The exact publication date, since Amazon's page shows none; we rely on Fortune.
  • Whether the memo's tone matches what managers actually do, which no source has measured.

What to watch next

Watch for Amazon engineers' own accounts of how the guidance changes reviews, for other large employers publishing similar guidance, and for whether the company replaces its flawed metrics with better ones, which the memo says it will work toward over time.

Sources

SaveStart the discussion

Written by

audiotale.net team

The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.

Read next