Harness as Code: Treating AI Workflows Like Infrastructure

Every AI coding session I've had follows the same arc: I open Claude/ChatGPT, explain what I want, go back and forth for 20 minutes, get something that mostly works, then close the tab. The code exists. The process that created it doesn't. Try reproducing that session tomorrow. Try reviewing the prompts your teammate used. Try version-controlling the "vibe" that made the AI generate good code that one time. You can't. And that's the problem. The IaC Parallel Remember when deploying infrastructure meant clicking through the AWS console? You'd configure a load balancer, set up security groups, tweak auto-scaling — all by hand. It worked, until someone asked "can you do that again exactly the same way?" and the answer was always "probably." Then Terraform happened. Infrastructure as Code. Define everything in .tf files, run plan, run apply. Reproducible. Versionable. Reviewable. The console still exists, but serious teams don't touch it. AI-assisted development is at the "clicking around