Amazon rethinks software engineering as AI takes on more coding

AI is changing how software gets built. At Amazon, that shift is prompting a broader review of engineering roles, team structures, and the way performance is measured.

According to AWS vice president and distinguished engineer Marc Brooker, AI agents now generate most implementation code for the vast majority of Amazon’s developers. The company has revised its internal engineering guidelines to reflect that change, emphasizing customer understanding, ownership, and business outcomes.

The implications extend beyond coding productivity. When implementation becomes faster, teams need to reconsider how they choose work, verify changes, and remain accountable for the systems they operate.

Measuring the value of the work

Amazon’s revised guidance encourages engineers to stay close to their customers and understand the products they build. An accompanying letter acknowledges that, as the organization has grown, some engineers and managers have become more removed from the business their software supports.

That distance matters. An AI agent can help implement a feature, but deciding whether the feature solves a worthwhile problem still requires context.

Brooker also cautioned against treating token consumption or lines of code as evidence of customer value. Amazon’s guidance allows a role for usage metrics, while emphasizing that they require careful interpretation.

The tension is visible in the company’s own experience. The Financial Times reported that Amazon discontinued an internal leaderboard ranking employees by their use of its Kiro coding tool after some employees began using it excessively to improve their position.

For engineering leaders, the practical question is what increased AI usage achieves. Faster implementation becomes valuable when it helps teams deliver useful changes, maintain reliability, or reduce the effort required to support a product.

Small teams, broader responsibilities

One example of Amazon’s changing approach is Velocity Labs, an AWS group formed to explore emerging AI ideas through rapid experimentation.

The team brings scientists and engineers together, encouraging both to work across their usual boundaries. Brooker attributes its speed partly to small teams, trust, and clear ownership: individuals take responsibility for part of a project and seek support when needed.

The group reportedly took Strands Decider 2B, a model designed to handle narrow decisions within AI agents, from concept to public release in less than a week. Its intended tasks include selecting tools and routing requests.

This illustrates how AI can shorten the path from an idea to a working release. It also highlights the importance of people who understand the underlying approach, can assess results, and know when an experiment is ready for wider use.

Production still requires engineering judgment

Rapid prototyping and dependable production software place different demands on a team.

A prototype can demonstrate that an approach is promising. A production system must also handle permissions, integrations, unexpected inputs, operational constraints, and future changes.

Amazon’s experience shows why those responsibilities remain important. The company acknowledged that an engineer following inaccurate AI advice contributed to a retail outage. In a separate AWS disruption discussed in the reporting, Amazon attributed the incident to a misconfigured role and said it introduced additional safeguards, including mandatory peer review for production access.

These incidents point to the need for controls around how generated changes reach live systems. Review, testing, access management, and recovery procedures remain part of delivery, regardless of who—or what—produces the implementation.

For software that supports daily business operations, the team must be able to explain how a change works, what it could affect, and how to recover if it fails.

Knowing when AI helps

Amazon’s guidance also recognizes that current AI tools will not improve every engineering task. Engineers who can explain where the tools fall short are exercising the judgment the company expects.

That principle gives teams room to choose an approach based on the problem. Some tasks benefit from an agent coordinating several tools. Others need a small script, a configuration change, or a clearer process.

Choosing well requires an understanding of complexity, operating costs, failure modes, and the effort needed to maintain the result.

What this means for software teams

Amazon’s approach reflects a wider shift in engineering: as code generation becomes easier, more attention moves toward problem definition, system design, verification, and operational ownership.

At Control F5 Software, those responsibilities connect planning, engineering, QA, and delivery. They matter when modernizing an existing platform, integrating AI into a workflow, or building software that a business will depend on over time.

AI can accelerate implementation. The lasting value comes from a team’s ability to turn that implementation into software people can rely on—and continue to improve.

Source

Control F5 Team
Blog Editor
OUR WORK
Case studies

We have helped 20+ companies in industries like Finance, Transportation, Health, Tourism, Events, Education, Sports.

READY TO DO THIS
Let’s build something together