My agents don't work in chat anymore. They work in files. I run 17 agents ever… · @ReignSwords · TwKit

@ReignSwords

My agents don't work in chat anymore. They work in files. I run 17 agents every night. They file 17 posts while I sleep. This is the loop that keeps them from eating each other — steal it: 1. One TODO.md per job. The agent picks the next item, does it, writes the result under that item, commits, moves on. No "done!" in chat. The file is the receipt. 2. Markdown beats chat history. Chat scrolls away. A file with validation notes and follow-ups is searchable in git forever. 3. Reviews become new TODOs. I never drop ad-hoc comments in chat. I batch remarks into the file as new items and rerun the loop. 4. One chat per topic. A new topic gets a new Markdown file and a forked chat. Mixed context is how agents start confidently doing the wrong thing. 5. Rules live at every level. AGENTS.md at the workspace level AND the repo level. The agent reads all of them. Generic model skill isn't enough — the local rules are the whole jump. The part nobody tells you: this loop also makes failures quiet. My free reviewer died silently at a 1,500-token cap and reported nothing. Another tool scanned a repo with a live AWS key inside and printed "✔ No suspicious files detected." 📌 Rule 6: no receipt, no done. Every job ends with one line — what it did, what it touched, what failed. If there's no receipt, it didn't happen. The model changes every month. The loop stays. That's the whole trick. Fellow AI builders — follow for more. My take: the agent isn't the product. The loop is. Fight me 👇 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥

曝光 69 · 评论 6 · 点赞 0 · 书签 0