How much can you delegate to agents?

The article argues that deciding how much autonomy to give AI agents depends not on model quality, but on two factors: how easy the agent's work is to verify and how costly mistakes are to undo. Based on these factors, tasks fall into four levels of delegation—from "agent as assistant" for hard-to-check, hard-to-undo work, to "self-driving mode" for easy-to-check, easy-to-undo tasks—providing a framework for safely maximizing agent autonomy in software development.