AI agents accelerate writing code, but they also speed up the spread of patterns in a codebase, any patterns.

When there are no clear rules, bad code and poor design choices spread quickly and become hard to fix. AI assistants cannot tell the difference between code that merely works and code that is well designed. This makes their uncontrolled use risky for large and evolving systems.

To reduce this risk, people must stay in the loop, but that is not enough on its own. Tests, code quality tools, and architectural checks are no longer optional. AI agents also need clear instructions on what maintainable, secure and performant code looks like.

This talk is based on a year of intensive, real-world use of AI agents by multiple teams and shares practical lessons on what worked, what failed, and what scales.

Eugene Fidelin is an Engineering Manager & Frontend SME @ eBay, ex Frontend Architect @ Marktplaats. Experienced Full-stack Web Developer with 20+ years in the field. Enjoy helping teams use AI agents to build well-tested, high-performing, and secure software using Javascript, Node.js and React.