Engineering Practices in the AI Coding Era

Software engineering responsibilities are reframed in the AI coding era. - AI lowers the cost of generating code; authors argue engineers must guarantee delivered code is verified. - Two required steps to prove code: manual testing (run and provide evidence like logs or screen recordings) and automated tests (submit new test cases that fail-before/pass-after). - Engineers should treat agents as junior staff: train agents to self-validate and write tests. - AmpCode experiment: codebases optimized for agents (letting AI choose naming/structure) improved agent efficiency; human renames can confuse agents and increase token use. - The concept of a codebase "By an Agent, For an Agent" emphasizes probability-consistent structure favoring AI maintainability even if it diverges from human stylistic preferences. - Human accountability remains essential: humans must certify correctness even when AI produces the code.
Key insights
Validate generated code via manual runs and automated tests
With AI-generated code becoming cheap, engineers’ primary responsibility is proving code quality. The recommended practice is twofold: manual testing—run code personally and include evidence such as terminal logs or UI recordings—and automated testing—add test cases that demonstrate the change (red-to-green). Authors emphasize combining manual and automated validation rather than relying solely on AI-generated outputs.
Treat agents as junior employees and build testing culture
Engineers should train AI agents to self-validate by instructing them to run code and write tests. A robust testing culture and well-structured test suites also enable agents to emulate existing patterns, increasing their usefulness and reducing risk of unverified code entering the codebase.
Agent-optimized codebases can increase AI productivity
An AmpCode case study found that allowing AI to control naming and file structure—rather than renaming to human-preferred conventions—improved agent efficiency. The result supports creating codebases that align with AI probabilistic expectations ("By an Agent, For an Agent"), which can trade some human readability for faster AI maintenance and fewer token costs.