Andrew Ng sorts AI coding work into three loops

Andrew Ng mapped AI product work onto three control loops: minutes, hours, and days, per The Batch. For a solo builder shipping with an agent, that split decides where your judgment has to insert. It also decides where you can let the process run without watching.

Three clocks, not one blurred workflow. That's Ng's point, and it tells you which clock a given decision actually belongs to.

The minutes loop is the agent writing and testing its own code, fast enough that you barely notice the wait. The hours loop is you, reading the diff and deciding if the agent's plan actually holds up. The days loop is the market: real users deciding whether what shipped was worth building in the first place.

The framework only helps if you know which loop you're in before you act. My read: solo builders default to babysitting the minutes loop because it's the one you can watch happen. They under-weight the days loop because it takes longer to resolve.