On September 4 the Next.js team described closing 1,500 GitHub issues in one month. An agent did the research across years of stale reports; every closure still required a human yes.
The pattern worth borrowing
- Research before reply: each report was checked against current behavior before anyone called it stale.
- Batch small decisions: evidence was pre-built, approvals ran in batches, not one spreadsheet at a time.
- Closed does not mean lost: still-real issues came back as reproducible reports instead of vanishing under a bot label.
If your own maintenance backlog has hundreds of dead tickets, this is the working recipe. The failure mode is its mirror image: an agent closing by pattern alone. The post is explicit that this one did not.