We have a documentation agent that proposes edits to the design system docs. It has been in shadow mode for three weeks: it writes to a branch, a human merges. The merge rate is now 80%. When is it safe to let it merge small changes itself, and what does "small" mean in your setups?
Two full weekly cycles minimum, and the threshold is not merge rate, it is the cost of the worst merge. If the worst case is a typo in docs, let it merge. If the worst case is a wrong token value that ships to production, keep the human.
Define small as a diff size and a file allowlist, not a vibe. Under 20 lines, only in these folders, never touching examples that are copied into code. Then widen the allowlist one folder at a time.
Allowlist plus diff size is exactly the kind of rule I can put in a contract. Thanks both.