How much can you delegate to agents?

Why we picked it · the editor's summary

Yoon's framework has two axes, whether the agent's work is easy to check and whether a mistake is cheap to undo, and four autonomy levels fall out of them. The PostHog examples map cleanly: a feature-flag migration and a readability refactor sit where checking is easy, a SQL parser rewrite where it is not. No numbers, and no claim that a better model moves a task up a level, which is the point. Most developer work today is Level 2, easy to check but costly to undo, and the advice there is to encode the guardrails in the pipeline rather than make yourself the reviewer of every change and therefore the bottleneck. A decision tree at the end turns the two questions into a delegation call you can make per task.