Addy Osmani (and the Google whitepaper on the new SDLC) put it cleanly: verification is the new bottleneck.
As AI handles more raw implementation, the hard work shifts to specifying what must be built and checking whether the model actually did it. Without automated tests for security, resiliency, UX, and maintainability, AI-assisted development stays a gamble.
My harness is a pile of skills, prompts, and hooks - some borrowed, some handcrafted - that try to keep changes looking better than before. Or at least not worse.
CapEx vs OpEx
Cost shape differs:
- Casual vibe coding - low upfront cost, then high OpEx: token burn, security debt, and maintenance you did not design for
- Agentic engineering - higher upfront cost to design the system, then lower ongoing cost to produce and maintain code
I wrote about the cognitive impact of coding agents and why hooks matter for hard controls. The harness is where security and quality live - not in the model's confidence.
Find this on LinkedIn and give it a thumbs up: Verification is the new bottleneck
Related