Coding agents need human oversight

This is a needed reality check on the current “delegate everything” hype cycle: speed is useful only when the team can still trust the codebase afterward.

• –Agents are most effective on bounded, testable tasks with clear evaluation loops; the article draws that line well.
• –The real risk is compounding small mistakes into duplicate abstractions, architectural drift, and a codebase no one fully understands.
• –Low-recall search in large repos means agents miss existing code and create more local optimizations than global coherence.
• –Human review is still the scarce resource for architecture, API shape, and release risk, so review bandwidth should set the pace.