AI Coding Made Us Faster. Then the Real SDLC Problems Started
AI coding is not another developer-productivity upgrade — it is an operating-model shock to software delivery, and we felt it within weeks of building our first real AI-native product with Claude Code and OpenAI Codex. The velocity problem gets solved; then every surrounding step — requirements, architecture review, testing, UAT, release governance, accountability — starts feeling the pressure.
Should we slow AI coding down? Probably not for long, but yes until the harness catches up. The deeper questions: where should human judgment sit when line-by-line review of AI output becomes either the bottleneck or a rubber stamp? (Answer: shift it left — requirement clarity, architecture direction, test intent, release risk.) Why is the PRD back? (As context infrastructure for agents, not bureaucracy.) And why does testing become the real control system when models can drop constraints as context shifts — or confidently defend an implementation that is wrong?
Adopting Claude Code was the easy part. Reorganizing the delivery organization around it is the real transformation.
Key Takeaways
- When one step gets 10× faster, every other step becomes the bottleneck
- Human-in-the-loop should move to the highest-leverage points, not the most familiar checkpoints
- The PRD returns — for the AI's benefit as much as the humans'
- Machines verify conformance to spec; humans validate that the spec is right
- Sprint mechanics need rethinking when dev capacity stops being the constraint