“Each person picks a main quest, as in World of Warcraft: the feature that moves the story, needs a whole model in your head and splits badly. Around it sit side quests (a fix, a doc, an upgrade) that coding agents run in parallel.”

Planning as a sorting problem: work that needs 'a whole model in your head and splits badly' stays with a human, and only the small independent stuff is parallelised. It pairs with his number from the same week: 948 merged PRs in four weeks, where the largest 10% (95 PRs) carry 71% of changed lines. That is the same shape: most agent PRs are small side quests, while a few big ones carry the real change. It gives a concrete heuristic for what not to hand to concurrent agents, which the log has little of on the planning side. Source is one post from a CPTO (reply in a thread), so it's a practice report, not a measurement of outcomes.