Running AI agents in production? You'll spend 80% of your time on operations, not prompting.
This repo is a living collection of battle-tested ops patterns — zero theory, exact commands, and a verification checklist with every playbook. Every pattern here was written from a real incident that cost time or money before we figured it out.
| # | Playbook | Problem it solves | New? |
|---|---|---|---|
| 1 | Post-Change Audit | AI modified the system — how do you verify it didn't break everything? | |
| 2 | Session Archive | Conversation history grows forever — archive without losing data | |
| 3 | Three Rules | Before parsing external formats or deploying cron — read the data first | |
| 4 | Self-Healing Watchdog | Who restarts the agents when every agent is down? | |
| 5 | Provider Fallback Chain | Your API provider goes down — 7 seconds of config keeps you running | 🆕 |
| 6 | Model Routing Diagnostics | Base URL mismatch? Session cache stale? One command finds out | 🆕 |
| 7 | WeChat Message Rules | Rate limits, retries, dedup — keep an agent alive on chat platforms | 🆕 |
"AI wrote the code. The watchdog caught the crash. The audit proved it was fixed."
这个仓库是 AI Agent 运维实战手册——所有内容来自真实线上事故。 你跑 agent 时遇到过的：API 挂了自动降级没生效？配置改了没重启？会话日志炸了磁盘？ 每个 playbook 都是一个具体问题的检查清单，复制粘贴就能用。
7 篇已发布，持续更新中。欢迎 Star ⭐
# Clueless about what's running on your system?
lsof -i -P -n | grep LISTEN
# Need to know if your agent actually restarted after a config change?
grep -c "listening.*port 8642" ~/.hermes/logs/gateway.log
# Cron job that should produce output but you never see it?
hermes cron run <job-id>If those commands gave you useful information, this repo is for you. Pick any playbook above and follow its checklist.
- Browse — each playbook is a standalone README with: problem → core insight → step-by-step flow → verification checklist
- Copy/paste — every command is copy-paste safe (tested on macOS/Linux)
- Contribute — hit a new pattern? Open an issue or PR. All patterns must include the real incident that produced them
This repo launched 8/13 with 4 playbooks. By 8/27 we've added 3 more from real production hits:
- Provider fallback chain — discovered when a $400 bill taught us the auto-fallback didn't exist
- Model routing diagnostics — discovered when 30 minutes of debugging was actually a one-command fix
- WeChat message rules — discovered when rate-limited delivery took out a week's worth of daily reports
The repo grows as operations happen. Star it if you want to watch it grow ⭐
- Verify, don't assume. Every claim ships with a reproduction command.
- Archive, don't delete. History is cheap; regret is expensive.
- Zero-LLM for mechanical work. Watchdogs and crons don't need a model.
- Post-change audit always. If you changed it, verify it — then record it.
MIT