From AI Coding to AI SDLC: What Four Iterations Taught Us
Four iterations into a real AI-first project, the early "AI can do everything" excitement has matured into a more interesting question: how do we build a better engineering system around AI? The journey moved through layers — first prompting (communicating intent, constraints and the why, not step-by-step instructions), then velocity problems (merge conflicts, branch drift, AI changes touching more files than expected), and finally the deeper realization.
AI agents are not gods. They are closer to very capable developers: fast, knowledgeable, always available — and also sometimes careless, too agreeable, forgetful, overconfident, or simply wrong. We spent decades building software engineering practices because human developers are imperfect; AI developers are imperfect in a different way. That is why the thinking shifts from "AI coding" to "AI SDLC": requirement grilling before implementation, spec-to-task workflows where task size and boundaries directly shape agent behaviour, TDD as a practical guardrail rather than a slogan, and LLM-based verification loops before human review.
Key Takeaways
- Treat AI agents as powerful new team members with new strengths and new failure modes
- Vague requirements make AI move fast in the wrong direction
- Task decomposition is now part of the SDLC itself
- Without strong validation loops, a model can anchor on a wrong assumption and build on it
- The winning teams engineer with AI properly — they don't just use it most