Jamon Holmgren wrote up his Night Shift workflow — a model where you do the thinking during the day (specs, architecture, requirements) and let AI agents execute autonomously overnight. By morning, you've got commits to review. He reports it's 5x faster with better quality than his previous approaches.
The bit that really landed for me: "You should be willing to burn all the tokens trying to make sure everything is as perfect as the agent can make it before a human ever has to review anything." Your time and energy are the expensive resource. Tokens are cheap. Act accordingly.
This is where my head's been for a while now. Why start from zero when you could start from a draft PR that just needs steering? Even when the agent gets it completely wrong, that failure is valuable — it tells you where your docs, specs, or tooling need improving. The throwaway work still compounds.
I'm moving more and more towards this default. Before I start anything, I'm asking "how can I automate this and then review it?" Imagine a ticket lands on your desk and by the time you're back from making a coffee, there's already a draft implementation waiting. Or three. All it needs is some correction. That's not laziness — it's leverage. I invest the same hours in my week, but the output scales because I'm reviewing and refining instead of creating from scratch.
Jamon's full thread on X is worth a read. Some of this connects to ideas I've been thinking about — like how subagents keep work in the smart zone, how context window management is everything, and how your codebase structure matters more than your prompt.