About 98% of agent failures raise no errors. The call succeeds, the logs look clean, and only the content is broken — that’s what we measured running 115 autonomous agents on local models, 24/7, on one machine, for a year (from 2,510 classified failure records).
This kit is only the holes we actually fell into. Not one item is a theoretical best practice.
What you get (free, by email)
- ✅ 24-point checklist — everything that breaks in unattended operation, each item dated to the real incident
- ✅ AGENTS.md template — an agent constitution with the safety rules included
- ✅ Fail-closed send gate (60 lines, zero deps) — “don’t send” as the default; born from two real mis-send incidents
- ✅ Missed-fire watcher (70 lines) — cron never tells you about the fires that didn’t happen
Get the kit
We’ll email you the download link. You’ll only ever receive measured findings about harness operation, and every mail has one-click unsubscribe.
Honest note
We have no customers yet. Everything here is exactly what we run on ourselves.
Operated by GX Cafe LLC, Japan. (特定商取引法に基づく表記 ／ プライバシーポリシー)
日本語要約
ローカルLLMエージェントの放置運用キット（無料）。115体・1年間の実運用で実際に踏んだ穴だけをまとめた24項目チェックリストと実戦テンプレート3本を、メール登録でお送りします。運営: GX Cafe合同会社