Amazon resets engineering standards as AI writes most code

By Tom Whitfield

More than 80% of Amazon developers are expected to use AI each week, according to The Financial Times, as the company resets engineering standards around a material change. Amazon says AI agents now generate almost all implementation code for the vast majority of developers.

Revised guidelines, sent earlier this year and accompanied by a letter published last week, shift assessment toward customer outcomes rather than token spending or lines committed. The stated policy says AI use does not directly enter reviews or promotions, though it may count when it strengthens performance; engineers must explain why a task does not benefit from AI.

What changes in daily engineering work?

Two years ago, most Amazon developers still wrote code by hand. Marc Brooker, an AWS vice president and distinguished engineer, said AI agents now generate almost all implementation code for most of that population, calling it “a massive change in the way of working for software engineers.”

Brooker now works with Kiro, Kiro Crew and Anthropic’s Claude Code rather than spending most time in a traditional editor or command line.

The accompanying letter, written by Michael “Terk” Terkowitz, an Amazon stores vice president and distinguished engineer, tells engineers to learn AI, remain close to customers and products, and recognizes that some engineers and managers became less connected to the business as Amazon grew.

Amazon had not previously published its underlying software-engineer role guidelines and declined to provide the revision. Brooker, a co-author of the letter, said the update emphasizes outcomes over inputs and described token counts and committed lines as industry anti-patterns that do not equal customer value. For Amazon, AI adoption must serve customer value, not become a proxy for effort.

How will Amazon measure performance without creating a token race?

The letter does not dismiss usage data. Amazon found that usage measures correlate with meaningful outputs and expects that pattern to continue for some time, while warning managers to use the data with care and judgment.

Earlier conduct created a risk of output gaming. Three developers told The New York Times last year that managers pushed them to use AI, raised output goals and became less forgiving about deadlines; some described their work as resembling warehouse work.

In May, Amazon shut down its KiroRank leaderboard after finding that some employees used the Kiro coding tool excessively to climb the rankings, The Financial Times reported. Brooker said he did not recall the episode or its details, so he could not say whether it prompted the revised guidelines.

The standard for refusal is explicit: “If your work does not benefit from current AI tooling, that is a finding, not a failing.” Engineers should document the reason, share it with peers, raise it with a manager and help find a solution. The policy works only if managers judge outcomes rather than usage.

What do AI failures mean for controls and review?

Amazon acknowledged in March that an engineer who followed inaccurate advice from an AI agent contributed to one of several outages on its retail site. It disputed a Financial Times report that junior and mid-level engineers would need senior sign-off on AI-assisted changes.

Amazon also rejected an earlier FT account of a 13-hour AWS disruption in December, when engineers allowed Kiro to change an environment and the tool chose to delete and recreate it. It blamed a misconfigured role and user error, and said it has since added mandatory peer review for production access.

Amazon is asking engineers to use AI freely while relying on human review where production access can turn a bad suggestion into an outage. Broad AI use still requires human controls around production access.

Can a small team turn AI coding into a faster product model?

Amazon formed Velocity Labs in April, and the roughly 35-person team is based mostly in Seattle. Its purpose is to build rapid prototypes in AI areas Amazon considers underexplored and important to customers, rather than follow the company’s usual product-oriented process.

Job listings say open-source peers and some competitors deliver ideas faster, so the unit must move faster, explore more and keep AWS at the cutting edge of agentic AI.

The unit deliberately mixes scientists and engineers, allowing each to work across the boundary between research and implementation. It sits within the agentic AI division led by Swami Sivasubramanian, which organizes work into teams of six to eight people, and includes the developers of Strands, AWS’s open-source agent-building toolkit.

Amazon released Strands Decider 2B with its training recipe on Oct. 1, and the team moved from concept to public release in less than a week. The open-source model, based on Alibaba’s Qwen, makes quick, narrow decisions within agents, such as choosing a tool or routing a request.

Brooker built the first version on a home GPU after TypeSafe AI released a similar Jev model on Sept. 15. He said every line was written by an agent, while its core ideas remained his or were understood by him. OpenAI announced a similar hosted Decisions API service that same week, according to TechCrunch.

Brooker credited small size and high trust: each person owns a project part and asks for help when needed. He expects Amazon to create more such teams, although not across the entire company. Small teams move fast, but speed alone does not prove safe scalability.

What does the workforce strategy mean for hiring and job security?

Employment signals conflict. Amazon confirmed roughly 30,000 corporate job cuts since late 2025, and CEO Andy Jassy expects AI to shrink the corporate workforce over time.

In April, AWS CEO Matt Garman said Amazon was hiring as many software developers as ever, while a spokesperson said more than 11,000 interns and early-career software engineers would join globally this year.

Brooker says AI does not change engineer numbers. Amazon wants engineers who take ownership, remain curious and learn quickly beyond their established expertise.

He rejects a universal AI prescription. An agent powered by a cutting-edge model can be the right choice, while three lines of Python can be equally suitable; the decision requires judgment, not ideological commitment. Amazon must turn output gains into better products and careers, not tighter quotas.