Ryan details how his team ran a five-month experiment building an internal beta product with over one million lines of code and thousands of PRs—entirely authored and reviewed by Codex agents. Central to this was 'harness engineering': shifting focus from prompting models harder to designing systems where agents operate autonomously—enabled by fast build loops (under one minute), observability-first tooling, structured markdown scaffolds (spec.md, quality score tables), and self-improving workflows via session log analysis. They developed Symphony, an Elixir-based orchestration layer for multi-agent coordination, and pioneered 'ghost libraries'—spec-driven, reproducible systems distributed as high-fidelity specifications. Humans moved from line-by-line review to systems design, encoding engineering taste and non-functional requirements directly into agent context. While current models still struggle with true zero-to-one product creation and deep refactorings, the frontier is clear: agent-legible software, token-efficient CLIs, enterprise-grade governance via Frontier, and a future where 'you can just build things' becomes operational reality—not marketing.