Amazon told its software engineers this week to stop chasing AI token counts. A new internal memo calls token counts, AI adoption rates, and lines of code "deeply flawed" ways to judge good work1. The same memo still tells managers to track those same three numbers, team by team.
Terk Terkowitz, an Amazon vice president and distinguished engineer, signed the memo. He calls this moment the era of the "Renaissance Developer"1.
The Renaissance Developer, Defined by Amazon
Terkowitz leads engineering guidance for Worldwide Amazon Stores. His memo argues that judgment, more than any single use of AI, now marks a good engineer.
"Our most prolific engineers have demonstrated that neither the developer role nor their output is defined by the use of AI," Terkowitz wrote1.
That line retires Amazon's old scoreboard. Engineers once climbed an internal leaderboard by burning through AI tokens. The industry calls that habit tokenmaxxing: running a coding assistant hard just to post a bigger number. Terkowitz wants engineers blending AI with their own skill instead, building things "neither humans nor AI tools could produce alone"1.
He frames the shift as continuity, pointing to a 2012 line from Jeff Bezos: "What's going to change in the next five to ten years? But I very rarely get asked, 'What's not going to change?'"1.
A Flawed Number Amazon Still Watches
Here is the twist Terkowitz's own memo prints. Token counts, AI adoption, and code-change volume stay "deeply flawed" measures of a team's worth, he writes. They drift from what actually matters to customers1.
Yet the memo tells managers to keep reading them. It recommends watching the 10th percentile of teams, the slowest to adopt AI. The 90th percentile, the heaviest AI users, gets the same scrutiny. Either way, the memo asks whether a team's output backs up its number1.
Terkowitz even credits the flawed metrics with a point in their favor: "these input metrics correlate well with the outputs that matter," he writes, "and we expect they will continue to do so for some time."
So Amazon kept its token counts, instead of dropping them. It reclassified them, from a target engineers chase to a measure managers weigh "with care and judgement"1. A team or an engineer who finds little use for AI gets cover from the memo, rather than blame.
"If your work does not benefit from current AI tooling, that is a finding, not a failing," Terkowitz wrote1. "An engineer who can explain where AI does not yet help their work is demonstrating the judgement we are asking for."
Why Outcomes Are Harder to Fake
A token count rewards volume. An outcome, in theory, rewards a shipped feature or a fixed bug, the kind of result a manager can check against a customer's experience. Amazon's cloud chief has been making that case since summer.
"I was talking to a room full of CIOs just a couple of months ago," Matt Garman, Amazon Web Services' chief executive, said. "I asked, 'How many of you are either seeing materially positive ROI today or have a path in the next couple of months to really high ROI?' 90% of hands went up, which is totally different than a year before"2.
Garman ties the shift to picking the tool that fits the job, over the flashiest one. "One of the things that's driven up a bunch of cost of AI is that people were trying to use the best model for every single thing," he said2.
Developers Use AI Daily, Trust It 25 Points Less
Share of developers who use AI coding tools daily versus who trust AI output, Stack Overflow's 2026 Developer Survey
Each column is a separate share of the same group of developers: the first use AI coding assistants or agents every day, the second trust what AI gives them only once they can check it themselves.
Source: Gergely Orosz on X [3]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Use AI daily | 73% |
| Trust once checked | 48% |
Outside Amazon, the data backs his instinct. Stack Overflow's 2026 Developer Survey found 73 percent of developers use AI coding assistants or agents daily. Only 48 percent trust the answers once they can check the work themselves4. Usage and confidence are two different numbers, and a token count only ever measures the first one.
69 Percent of Developers Point AI at Code They Know
Share of developers who use AI for each task, Stack Overflow's 2026 Developer Survey
Each bar is the share of developers who turn to AI for that task. The longest bar is also the task where a developer already knows what the right answer looks like.
Source: Gergely Orosz on X [3]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Code they already know | 69% |
| Debugging and refactoring | 64% |
| Writing tests | 58% |
That same survey shows where developers actually point their AI tools. Sixty-nine percent use them on code they already know. Debugging and refactoring draws 64 percent, and writing tests draws 58 percent4.
Developers lean on AI hardest in the places they can still check its work by hand.
Trillions of dollars sit behind this question, well past Amazon's own org chart. Gartner expects worldwide IT spending to reach $7.35 trillion in 2027, up 15 percent from this year. Data center and AI software spending drive most of the new money5.
"Collectively, the market forces are reinforcing spend around AI infrastructure, AI-enabled cloud platforms and advanced software ecosystems," said John-David Lovelock, a distinguished vice president analyst at Gartner. Real money now rides on whether any company can tell a busy AI dashboard from a useful one.
The Divide Gergely Orosz Sees Coming
Gergely Orosz edits The Pragmatic Engineer, a newsletter read widely inside the industry Amazon's memo addresses. He read the memo the morning it reached the public and posted his own verdict within hours.
"Amazon's leadership is very good at writing, but you still need to read between the lines," Orosz wrote. "What I read between these lines: Engineers who are not 'product engineers' (aka eng + PM interest) and don't like to be customer-facing will not have a future inside of Amazon"3.
Gergely Orosz@GergelyOroszAmazon's leadership is very good at writing, but you still need to read between the lines.
What I read between these lines:
Engineers who are not "product engineers" (aka eng + PM interest) and don't like to be customer-facing will not have a future inside of Amazon
Full:… pic.twitter.com/exf4g6z99U
Orosz is reading the same document Terkowitz signed, and finding a narrower message inside the warm one. Amazon's memo answers part of that reading directly. A team or an engineer who gets little lift from AI tooling earns a finding, the memo says, rather than a failing1.
It stays silent on an engineer who simply prefers writing code alone, away from customers. That is the exact profile Orosz says the memo is quietly sorting out.
Both readings can sit inside the same document. A company can mean its warning against tokenmaxxing honestly. It can still be building, sentence by sentence, a picture of the engineer it wants to keep.
What to do
If you manage people who use AI tools, read your own usage dashboard the way Amazon just read its own. A high count of prompts or tokens shows someone stayed busy. The quality of their work remains a separate question.
Ask instead for the outcome. A closed ticket counts. So does a fixed bug, or a customer who got a real answer.
Then ask the engineer to explain, in their own words, where AI fell short. Amazon now treats that answer as a sign of judgment instead of a confession.
By the numbers
- Nine in 10 CIOs told Matt Garman they see positive ROI from AI, up sharply from a year earlier2.
- 73 percent of developers use AI coding assistants or agents every day, Stack Overflow's 2026 survey found4.
- Fewer than half, 48 percent, trust AI's answers only once they can check the work themselves4.
- Most AI use stays close to home: 69 percent of developers point it at code they already know how to write4.
- Two percentiles, the 10th and the 90th, are what Amazon uses to sort teams by AI use, the metric its memo calls "deeply flawed"1.
- $7.35 trillion: worldwide IT spending Gartner forecasts for 2027, up 15 percent from 20265.
What to watch
Watch whether Amazon's performance reviews actually change, beyond the memo's own words. A manager who still ranks a team by its percentile of AI use, quietly, would answer Orosz's reading before Amazon's own words do.
Other large employers may publish their own version of this memo next. Amazon rarely moves alone on how it measures engineers. A second company's language will show whether "outcomes over tokens" is a lasting standard, or one memo's phrase.
Sources
- Terk Terkowitz, "An update from Amazon engineering leadership on the role of the developer in the age of AI," Amazon, Oct. 27, 2026, https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai
- Amazon Staff, "AWS CEO Matt Garman on why enterprise AI is finally delivering real returns," Amazon, Aug. 1, 2026, https://www.aboutamazon.com/news/aws/aws/aws-matt-garman-enterprise-ai-roi-interview-2026
- Gergely Orosz, "Amazon's leadership is very good at writing, but you still need to read between the lines," X, Oct. 28, 2026, https://x.com/GergelyOrosz/status/2115331737714622800
- Stack Overflow, "2026 Developer Survey: AI," Stack Overflow, Oct. 6, 2026, https://survey.stackoverflow.co/2026/ai
- Gartner Staff, "Gartner Forecasts Worldwide IT Spending to Grow 15% in 2027, Totaling $7.35 Trillion," Gartner, Oct. 21, 2026, https://www.gartner.com/en/newsroom/press-releases/2026-10-21-gartner-forecasts-worldwide-it-spending-to-grow-15-percent-in-2027-totaling-7-point-35-trillion