Hermes doesn’t fail on capability — it fails on setup habits. These are the ten Hermes agent best practices I actually run daily, in the order that matters, from memory-first to the budget caps that stop an unattended agent burning money overnight.
Short answer
- Memory first, persona second, interfaces last — most people do it backwards.
- A profile per job and per model; test tool calls, never chat.
- Approvals on, /journey audits weekly, exports weekly.
- Match models to jobs, and cap anything that runs unattended.
The practices that separate a working agent from an abandoned one
The failure pattern with Hermes is never capability. Someone installs it, it’s impressive for a day, then every session starts blank, every job needs babysitting, and a week later it’s abandoned.
Every practice below fixes a specific version of that failure, and every one comes from running Hermes daily in a real business — not from a docs page.
1. Build memory before anything else
An agent with no memory starts every session like a first date — you re-explain yourself, your business and your projects hundreds of times a year, and the output comes back generic.
The fix is a vault of plain markdown your agents read before working and write to after finishing. Your business, your clients, your decisions, your voice. And crucially, your agents maintain it — you don’t write markdown, they do. Full setup in the memory guide.
2. Get the persona right second
The SOUL.md file defines who your agent is — how it talks, what it cares about, what it’s allowed to do. It shapes every single response, which makes it the highest-leverage file per minute you spend on it.
Most people do this backwards: a beautiful dashboard in front of an agent that knows nothing about them. Memory first, persona second, interfaces last.
🔥 Want this set up without the guesswork? These practices are exactly what the 30-day roadmap walks you through in order. Inside the AI Profit Boardroom you get the Agent OS as a ready-to-install file, a 30-day roadmap, daily tutorials the same day new tools ship, and four live coaching calls a week where you share your screen and get unstuck. → Get access here
3. Run a profile per job, and per model
A profile is a fully isolated setup — its own config, memory, skills, credentials and history. Use that.
- One profile per job: research, content, inbox, outreach.
- One profile per API you’re testing, so you can run two agents on the same task side by side and compare honestly.
- A new model becomes a new profile, not a migration.
4. Test tool calls, not chat
Plenty of setups answer questions beautifully and fall over the moment the agent has to actually do something. Before trusting any new model or profile, ask it to run a skill, read a file, search the web. That’s the only test that tells you anything.
5. Keep approvals on
Hermes runs real commands on your real machine. There’s a mode that skips confirmation prompts and it feels fast — leave it alone while you’re learning, and turn it off later on purpose, when you know exactly what you’re allowing.
The same caution scales up: an email agent gets a separate inbox, never your main one, and says plainly that it’s an AI agent. Anything browsing live pages is reading content you don’t control.
6. Watch the context meter, and start fresh
There’s a live meter showing how full your session’s context is — click it and you see what’s eating the space. Long messy sessions are where quality quietly drops. When it’s filling up, start a fresh chat instead of pushing on.
7. Audit what it’s learned, weekly
Run /journey and you get a timeline of every skill and memory your agent has built up — scrub through, edit or delete anything wrong.
This matters because self-improvement compounds whatever works, including mistakes. An agent that learned something wrong will keep being confidently wrong until you look. Clean memory means a better agent.
8. Export weekly
One command packages your entire agent — persona, skills, memory, cron jobs, plugins, settings — into a single file, with credentials stripped automatically. Import brings it back on any machine.
Weeks of tuning shouldn’t live on one laptop. Export before every experiment and every rebuild. Details in backup and restore.
9. Match the model to the job
| Work | Model tier |
|---|---|
| Planning, hard reasoning, final review | A frontier model — runs once |
| High-volume routine agent work | A cheap model with strong caching |
| Tool calls, memory lookups, sub-agent grunt work | A small model like LFM2.5-2.6B |
| Anything confidential | A local model — nothing leaves your machine |
Sending every task to your most expensive model is sending your best person to sort the post.
10. Pick the right automation layer — and cap it
/loop is timer-driven polling inside a session. /goal iterates until an objective is judged done. Cron runs unattended, outside any session. Picking the wrong one is the most common automation mistake — the split is in the loop guide.
And anything unattended needs a ceiling. Every loop tick is a full agent turn, and a scheduled agent pointed at a paid API can run up a bill overnight in a way an interactive session never will. Budget caps are not optional.
Want these practices as a system instead of a checklist? The Agent OS in the AI Profit Boardroom ships with the memory vault, profiles, skills and scheduling wired in — install file, 30-day roadmap, daily tutorials and four coaching calls a week. Start free with the free AI course and community.
FAQ
What are the most important Hermes agent best practices?
Memory first, persona second, a profile per job, tool-call testing, approvals on, weekly /journey audits, weekly exports, model-to-job matching and budget caps on anything unattended.
What should I set up first?
The memory vault. It’s the piece that can’t be rebuilt from scratch and the reason output stops sounding generic.
Should I turn approvals off?
Not while learning. Hermes runs real commands with your permissions. Turn them off later, deliberately, once you know exactly what you’re allowing.
How do I stop my agent learning bad habits?
Check /journey weekly and edit or delete anything wrong. Self-improvement compounds mistakes as happily as wins.
How often should I export?
Weekly, and before any rebuild or experiment. It’s one command and it’s the difference between a bad afternoon and losing weeks of tuning.
One profile or many?
Many — per job and per model. Isolated memory keeps experiments from polluting your working agent, and comparisons stay honest.
How do I keep costs down?
Route by difficulty, use caching-friendly and free models for volume, and put hard caps on anything that runs unattended.
Loop, goal or cron?
Loop for polling in a session, goal for one objective iterated to done, cron for unattended schedules. Wrong layer, wasted tokens.
The bottom line
Hermes agent best practices come down to sequence and caution: memory then persona then everything else, a profile per job, tool calls as the real test, approvals on, weekly audits and exports, models matched to work, and a cap on anything that runs while you sleep. Do those and the agent you have in month three is one nobody starting fresh can catch.
About Julian Goldie
I run Goldie Agency, a 7-figure SEO agency, and teach this daily on a 400K+ subscriber YouTube channel. 240+ client projects on Upwork at a 100% job-success score, 10+ years through every major Google update. My systems are in the AI Profit Boardroom; my link building book is free here.
Related reading
Last updated August 2026. This is the living guide to Hermes agent best practices — it gets updated as the tools change.
