
My Hermes Agent Works All Night. Here’s What I Woke Up To.
Six weeks ago I set up Hermes Agent on my home server and pointed it at my Obsidian vault. I thought I was building a chatbot with file access. What I actually built was something that keeps working long after I walk away from the keyboard. The YanXbt article “Hermes Agent Builds Itself While You Sleep” went viral in the community for a reason — it describes a pattern that sounds theoretical until you live it. I am living it.
From One Cron Job to Forty-Two
It started small. One cron job that fetched AI news at 7:30 AM and dropped it into a WordPress draft. Harmless. Useful. I added another for EU AI news at 8:00. Then a daily AI image generator at 8:30. Then a vault content creator at 10:00. Before I knew it, I had 42 scheduled jobs running every day — health monitors, backup scripts, research aggregators, clipping processors, blog publishers, night synthesis engines.
The weird part? I never consciously decided to build a 42-job system. Each one made sense in isolation. One morning I realised I had an employee that never slept, never complained, and worked for pennies. That morning I also realised I had no idea what it was actually doing at 3 AM.
The Compounding Effect Is Real
YanXbt mapped out a growth curve: day one you have 3 built-in skills and an empty memory. Week two you have 12 self-created skills and 50 wiki entries. Month one you have 30 skills and the agent starts suggesting patterns you missed. I can confirm: that curve is conservative. My vault now holds months of processed clippings, interconnected wiki pages, nightly syntheses that connect ideas across topics, and a CRM that grows with every conversation.
The key insight YanXbt got right: Hermes does not reset between sessions. Every night builds on the previous one. The vault grows. The memory consolidates. The skills improve. There is no ceiling because the agent keeps its context across days — not just within one chat window.
What I Learned By Looking at What Happens at Night
Three things surprised me. First: the self-improvement loop works. The agent creates skills on its own now — small reusable modules that handle specific vault tasks. Second: overnight work surfaces blind spots. My night synthesis cron started connecting ideas I had filed in separate wiki pages. I discovered I was building a protocol stack (MCP, A2A, ACP, Hermes-otel) without realising it until the nightly reports showed me the pattern. Third: 42 jobs need monitoring. Not because they fail — they rarely do — but because you want to know what your agent learned while you were offline.
That last point is the one nobody warns you about. You are not building a tool. You are building something that grows. And growth without visibility is a gamble.
Practical Steps If You Want This Too
Start with one cron job. Not ten, not five. One. Let that run for three days. Check the logs. See what the agent actually produced. Then add one more. The YanXbt sequence is the right one: messaging first, then one cron, then verification, then more crons. I added a five-minute morning review routine — kanban, cron check, skills review, wiki scan, usage check. It takes seven minutes and tells me exactly what happened overnight.
The best part? I get a briefing every morning that references my projects, my vault, my notes — not generic AI slop. That changes how you start your day.
What is the first cron job you would set up if you knew your agent would keep working while you sleep? I would love to hear your setup. Comments are open.
Wat vond je van dit bericht?