Can AI realistically replace human developers for production code?

If you’ve shipped features with AI assistants or agents, how close are they to end-to-end delivery without human intervention-requirements clarification, architecture, tests, debugging, security, and long-term maintenance? Where does AI excel vs. fall short (legacy code, ambiguous specs, edge cases, performance, compliance)? Please share concrete examples, tools used, and metrics (bug rates, cycle time, MTTR).