Everyone is selling context engineering as the new discipline. On my machine it's plumbing: four boring pipes that feed the agent, one drain that keeps junk out, and a valve that decides which pipe answers which question.
A static Astro site with 15KB of JavaScript was shipping 3.4MB of images. Here are the three failure modes that got past the image pipeline — the fallback-always-ships rule, the screenshot the optimizer never touched, and the boilerplate nobody referenced — and the 15 lines of Sharp that fixed all of it.
My cleanup list said eight branches were unmerged. Every single one had already shipped to production. The problem wasn't the branches. It was the question I was asking git.
No server does not mean nothing to configure. These are the four response headers I added to my static site, what each one blocks, and the one I left out on purpose.
Today a second AI agent started up in a directory where another session was mid-work on a dirty branch. It was one git switch away from committing onto the wrong branch. Here's the boring, native fix.
The honest version of overnight AI development: the agent opens pull requests, a human merges. What ships by morning is reviewable drafts and done-when-green fixes — never an autonomous push to prod. Here's the setup that makes it real leverage instead of a liability.