A concrete approach to get 2x–4x engineering productivity with AI-assisted development—without sacrificing quality, reliability, security, or compliance. No vibe coding; no AI slop. You write specs with an agent, the agent implements, you review and refine, followed by agent and human PR review until the work is production-ready.
For engineering leaders and practitioners who ship real software and want to adopt AI tools in a disciplined way. The full doc covers:
- Before you turn the agents loose — What “real software” means in this context, making AI available and mandatory, guardrails, and the non-negotiables
- The SAW workflow — Collaborative spec → agent implementation → review (code, tests, pre-commit) → PR → agent PR review → address feedback → human review
- Process variations — Research, multi-phase work, specs as artifacts
- Tooling — AGENTS.md, skills, MCP servers, sub-agents (when and how to use them, what to watch out for)
- Beyond the core workflow — Expanding the AI boundary, safety and sandboxing, leveraging juniors, agent diversity, adapting for your use
- Appendix — Worked examples with real artifacts and PRs