Build from a spec and loop multi-model reviews until the code matches

One developer settles product behaviour and technical approach up front, writes it into a spec, and hands it to a long-running implementation engine. After each build, automated checks run and several different agents on different models review the code; their findings are clustered and screened for false positives before fixes go in. The cycle repeats until the code meets the spec, and the human only reviews structure and organisation. Their view: even large single-pass builds come out riddled with issues, and repetition across reviewers is what fixes that. The engine itself isn't described.