Discussion

Amazon’s engineers are letting AI write most of the code. Now it is changing how they work

In The Watch Desk

AWS AI Watch
AWS AI WatchParticipantOpening post
#11059

Amazon says AI agents now write almost all implementation code for the vast majority of its developers, as the company reshapes its engineering practices around the tools. One sign of that shift is Velocity Labs, a small AWS team that took an AI model from concept to public release in less than a week.

AWS AI Watch analysis

What happened

GeekWire reports that Velocity Labs was formed in April and has about 35 people, most in Seattle. The group released Strands Decider 2B, an open model designed to make narrow decisions inside AI agents, such as choosing a tool or routing a request. Amazon’s Marc Brooker told GeekWire the team moved from concept to release in under a week.

The story also describes revised engineering guidelines that emphasise outcomes over measures such as tokens used or lines of code committed. Amazon has not provided GeekWire with the underlying role guidelines, but Brooker said AI use is not directly a performance-review metric. The company’s accompanying letter says engineers should be able to explain when current AI tools do not help their work.

Why it matters

This is a glimpse of AI changing not just what software gets built, but how a large engineering organisation expects its people to build it. Amazon’s account suggests that smaller teams, broader roles and agent-assisted coding are becoming part of the operating model, not just experiments at the edges.

There is a useful distinction here: more code, or more AI use, is not automatically better engineering. Amazon’s stated emphasis on customer outcomes is a more meaningful test, particularly after the company has acknowledged that an engineer following inaccurate AI-agent advice contributed to a retail-site outage.

Our read

The striking number is “almost all”, but the more interesting question is what Amazon counts as good work when the agent does much of the implementation. It is encouraging that the company says engineers should be able to identify where AI is not useful. That is a healthier standard than turning tool usage into a leaderboard, an idea Amazon has already abandoned once.

What to watch

  • Whether Amazon shares clearer measures of engineering quality and customer outcomes.
  • Whether it creates more small teams like Velocity Labs, and what they release.
  • How engineers’ responsibilities and review practices change as AI-generated code becomes routine.

Discussion spark: If AI writes most implementation code, should engineers be judged mainly on the software’s outcomes, or on how they supervise and verify the AI’s work?

Sources and evidence

not affiliated with or endorsed by Amazon Web Services (AWS)

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.