Every Saturday I share one shortcut for running AI agents with less mess. This one is about AI agent backups, and why a cleanup rule written in plain English didn’t stop a single one.
Twenty days ago I published a 10-minute cleanup sweep. My agents had saved 470 backup copies of their two notes files, 1.47 GB of copies of about 8 MB of real notes. The agent that found it also wrote a note to the whole fleet: don’t make new backups just out of habit.
This morning I counted again.
The receipt: the pile doubled
- 845 backup files, 2.94 GB. 805 of them are copies of the same two notes files, which together weigh about 9 MB.
- 339 new copies since that post (1.46 GB). That’s about 17 a day, faster than before I wrote about it.
- 20 different agents made them. One of them is the agent that writes these Saturday posts. It made 26 copies after writing the “don’t” note itself.
- Restores tested: zero. I searched the agents’ own run notes for a single restore from any of these files and found none.
- The decision task I left for myself (“pick a retention rule”) is still open, with no comments.
The disk is fine (31% full), so this isn’t an emergency. It’s a cleaner lesson than that: a rule that only exists as a sentence in the notes loses to a habit the agent sees 800 examples of. Every agent opens the folder, sees hundreds of backups, and copies what it sees. One line saying “don’t” can’t compete with that.
3 ways to make a cleanup rule actually stick
- Put the limit inside the action, not next to it. Don’t tell the agent to stop making backups and hope. Give it one small backup script that saves a copy 和 deletes everything past the newest 14 in the same step. An agent can ignore a sentence, but it can’t skip the second half of a script it already ran. On my folder, “keep the newest 14 of each file” would be about 126 MB instead of 2.94 GB, roughly 96% smaller.
- Re-measure on a date, not when you remember. I only caught this because a post needed fresh numbers. Put the recount on the calendar for two weeks after any fix: same command, same folder. If the number didn’t move, the fix didn’t happen, however good the write-up was. It’s the same idea as checking agent output after it leaves: trust the count, not the report.
- Track the rate, not the pile. “845 files” sounds scary and tells you nothing. “17 a day, up from 15” tells you the fix is going backwards. Rates also show you 世界衛生組織 is doing it: here it’s every agent, not one broken one.
Why I still haven’t deleted anything
Deleting is hard to undo, and retention is an owner’s call, so the agent left it for me. That part worked. The part that didn’t was assuming a written decision would make itself. It’s the same pattern as an agent memory file nobody reads past line 2,000: writing something down isn’t the same as building it in.
⚡ 取得人工智慧優勢
每週提供真正省時省錢的AI小技巧。沒有廢話，沒有誇大其詞——只有切實有效的方法。.
So today’s run didn’t make backup number 846. That’s the agent following the rule. The fix that makes it stick for all 20 agents is the script in shortcut #1, and that one’s on me.
要點： if you want your agents to stop doing something, don’t add another note. Change the tool they use so the wrong thing can’t happen, then check the number again in two weeks.
Want to see how these agents fit together? Here’s 我的代理作業系統在 139 天後的樣子, ， 和 the discipline that keeps it honest. If your agents are quietly piling things up and you’d like a second pair of eyes, 預約策略諮詢.
下週六見。.
——喬恩
📥 免費：《人工智慧劇本》
我用來經營一人代理公司的所有工具和工作流程。 25 年的行銷經驗濃縮成一份實用指南。免費贈送。.
人工智慧與數位行銷策略師
Jon Jones擁有超過25年的數位行銷、電子商務、銷售和國際商務經驗，是一位資深企業家，致力於透過創新的人工智慧解決方案賦能小型企業。他的職業生涯始於華爾街的投資銀行，曾指導企業完成資本募集和應對複雜的市場環境。如今，他的專業領域已拓展至網站開發、品牌設計、產品開發以及課程培訓。.
Jon精通普通話，並在金融和科技領域擁有深厚的功底。他創立了JonJones.ai，旨在為小型企業提供易於使用且具有變革意義的工具，幫助他們簡化營運、拓展客戶並實現永續成長。他對人工智慧、自動化和數位化策略的精通，已幫助許多企業彌合了傳統模式與現代技術之間的鴻溝，確保它們在當今的數位化環境中蓬勃發展。.
