We were taught to build to last, to build reusable stuff, to preserve, to standardize. Now think of scaffolding - it checks all the above boxes, but it allows you to assemble it differently on every single building. Now that producing source code, and thus assembling components, is cheap - why think of your most internal SDLC loop as something rock solid, if your development concept can be kept rock solid while the actual development loop can be adapted to a specific task? The same question goes for the AI harness - why keep struggling with a harness which keeps changing and which provides you with a zillion capabilities you do not actually need for a specific task? In this talk, I'll take you through the process of building disposable niche, but deterministic harnesses which leverage stochastic AI processes where they have to - as opposed to using stochastic AI harnesses and hoping they'll actually use the deterministic tools and gates you told them to.