用 Grok Bot 第一週學到的 3 件事，讓用量撐得更久……

用 Grok Bot 第一週學到的 3 件事，讓用量撐得更久。 用量提早見底，不是聊太多。是有一份工作一直在跑。 我跑開發名單迴圈，工具用量很高，很早就撞上限，把數據拆開看。真正要緊的是： 1. 小批處理，然後停。 不要給 agent 一個沒邊界的目標。做完一小段、存結果，再繼續。一次「做到完」的長跑，可能比一早上受控的工作更燒。 2. 只把入圍名單往下傳。 前面步驟可以看很多。後面步驟只該看到過篩的。把整份工作集往後帶，後面每一步都更貴。 3. 給迴圈硬上限。 最多幾筆、最多幾頁、最多重試幾次。你不叫停，agent 就會一直點、展開、呼叫工具。 Agent 計的是步驟和 token，不是訊息。 把工作設計成能在一次短跑做完，不然用量會過早見底。

原文：The 3 lessons I learned from my first week using Grok Bot to make my usage limits last. Reaching your usage limits early isn’t from chatting too much. It’s from one job that keeps running. I was running lead-generation loops with very high tool usage, hit my limits early, and dissected the data. Here’s what mattered: 1. Batch small, then stop. Don’t give the agent one open-ended goal. Finish a short slice, save the result, then continue. One long “keep going until done” run can burn more than a whole morning of controlled jobs. 2. Only pass the shortlist downstream. Early steps can look at a lot. Later steps should only see what survived the filter. Carrying the full working set forward makes every following step more expensive. 3. Put hard limits on the loop. Max items, max pages, max retries. If you don’t stop it, the agent will keep clicking, expanding, and tool-calling. Agents bill steps and tokens — not messages. Design the job so it can finish in one short run, or you will hit your usage limits way too early.

整合