AI Can Write More of the Code. We Still Have to Engineer the System

AI can generate more implementation code and assist with analysis, design, and testing, but that does not transfer responsibility for choosing what to build or deciding whether the result is acceptable. Engineering judgment remains essential: decisions must stay visible beyond a model conversation, context must be selected deliberately, and generated changes need evidence before approval. A concise project guide can record scope, system boundaries, architectural reasoning, quality expectations, and verification commands. For each increment, engineers should provide relevant context and acceptance cases, then check that the agent has understood the request before it implements anything.

A hypothetical mobile search feature illustrates the approach: define local search behavior and exclusions, check matching and empty results, and prevent stale asynchronous results from replacing newer ones. Inspect the diff, run focused checks, and exercise the interface; formatting, compilation, tests, and screenshots each reveal different things, and none proves that the feature solves the right problem. Documentation should be proportional to risk: detailed specifications may be warranted for safety, regulation, compatibility, or coordination, while redundant documents create review and maintenance costs. The proposed cycle - specify behavior, supply context, verify interpretation, inspect changes, gather evidence, and make an explicit acceptance decision - is a practical position, not a demonstrated productivity gain. Any claim of improved efficiency would need to count review, rework, defects, and maintenance, not just code-generation time.