“From Sept 22 to Oct 20, we merged 948 pull requests in our main repository. The largest 95, or 10%, carry 71% of all changed lines. The other 853 share the remaining 29%.”

Counting PRs hides where the load is. A reviewer or merge queue facing 948 PRs a month really faces ~95 big ones plus a long tail of small ones. That argues for routing the two classes differently (light review or auto-merge for the tail, human attention for the head). The post doesn't say how many PRs were agent-authored, so treat it as context for the planning split in the follow-up reply, not as an agent-only statistic.