我今天试着让 claude code 启用 workflow 功能，它开启了 273 个 Agent 来帮我总结 1 万个客服会话。
然后，启动了半小时，我问了它一句：
我想问一下，workflow 使用的 model 是什么？可以调整成 sonnet 做摘要总结吗？
它回答：
Good question — let me answer precisely and check the current run’s progress to decide the smartest switch.
What model the workflow uses: the agents inherit the main-loop (session) model because I didn’t set a per-agent model override. Your session is currently Opus 4.8, so the running summarization is using Opus 4.8. Yes — I can pin the agents to Sonnet via opts.model: ‘sonnet’, which for ≤100-char summaries + classification is the right tool (much cheaper/faster, quality is plenty).
Let me check how far the Opus run got so I can switch without wasting completed work:
91/273 done (~33%) with Opus, ~9,100 sessions remaining. Switching the remainder to Sonnet saves real cost, and the 91 completed Opus batches are good — I’ll keep them. Let me stop the Opus run, bank the completed work, then run the rest on Sonnet.
嗯，反正我反应过来的时候，5 小时用量已经打没了。
希望对你有用。