9 Hermes Agent Lessons From 100+ Hours: Models, Failover, Tailscale, Crons

Step-by-Step: The 9 Lessons

1. Pick your model by budget.

   Opus is the most reliable — it "crawls to the finish line" and completes tasks no matter what (he spends ~$40/day / ~$1,400/month on Opus API for Hermes alone). ChatGPT 5.5+ is a usable, cheaper option if you already subscribe (anything before 5.5 was not usable). GLM 5.2 is the ultra-cheap pick — a bit robotic but it gets the job done. For a serious revenue business, he recommends Opus for peace of mind.

2. Run at least two Hermes agents.

   Keep a second profile on a different model (e.g. an Opus "Hermes" main plus a ChatGPT-backed "GPTme's" backup) so they monitor and repair each other. Hermes still breaks occasionally — multiple agents give you failover.

3. Don't over-isolate accounts.

   You don't need a separate Mac mini, Gmail, or Apple ID for every agent. Hermes only does what you tell it — it won't rifle through your photos or email your contacts unprompted. Over-isolation just adds friction; he runs it on his main machine and main accounts. (See our security caveat below — scope this to your own risk tolerance.)

4. Use the right platform for the moment.

   Hermes desktop at your computer (switch profiles, see cron jobs, pin/pop-out sessions, talk to multiple agents at once); Telegram for deep work on the go; iMessage for quick prompts (add a "Hermes" contact and pin it to the top).

5. Exploit Telegram formatting.

   Recent Hermes updates render tables, bold, and paragraphs in Telegram. Schedule cron jobs that return formatted tables — e.g. a daily AI-stock research table you wake up to.

6. Clean up cron jobs to fix a slow agent.

   Background crons are the #1 cause of a sluggish, "always busy," or "getting stupid" agent. In the desktop app's cron list, pause the ones you no longer use every week for an instant performance boost and token savings.

7. Install Tailscale (free) for multi-device.

   It creates a private network so a single Hermes agent can SSH across all your machines — check local models on a DGX Spark, move files between computers, or reach a localhost dev server from your phone. This is where Hermes becomes a real "AI employee" running your whole fleet.

8. Reverse-prompt every morning.

   Run a prompt that has Hermes interview you about the day's priorities, tasks, and stresses, then propose what it can take off your plate and automate. It surfaces 2–3 new jobs a day in ~5 minutes.

9. Push those tasks to the Kanban board.

   Open the Hermes dashboard, go to the Kanban board, drop in the tasks from the morning interview, assign them to the agent, and let it work them — a clean way to stay organized.

Commands & Code Shown

hermes-dashboard

hermes-dashboard

Purpose: Opens the Hermes web dashboard (profiles, cron-job list, and the Kanban board) in your browser.

When to use: Managing profiles and crons, and dropping morning-interview tasks onto the Kanban board. Run it from Ghostty or your built-in terminal. See our Hermes dashboard guide.

Create a second profile (natural language)

Set me up a new Hermes profile powered by ChatGPT. Name it GPTme's.

Purpose: Spins up a second Hermes agent on a different model for failover. ("Profiles" are Hermes' term for separate agents.)

When to use: Any time you want a backup agent — or use the dashboard: Profiles → Create.

Reach another device over Tailscale (natural language)

Go to my DGX Spark and list which local models are running.

Purpose: With all devices on the same Tailscale network, the agent SSHes across machines to run commands, check models, or move files.

When to use: Running a headless device fleet (Mac Studio + DGX Spark + laptops) without plugging each into a monitor.

Common Errors & Fixes Covered

Gotchas & Caveats

• The "don't over-isolate" advice reflects his personal risk tolerance. For agents that touch email, messages, payments, or sensitive files, scope permissions tightly and consider a dedicated account — see our Hermes security guide. An agent doing exactly what you tell it can still be a problem if you tell it the wrong thing, or if a prompt-injection source is in the loop.
• Opus via API is expensive (~$1,400/month in his case). GLM 5.2 or an existing ChatGPT subscription are far cheaper if you can accept slightly lower reliability.
• Tailscale's free tier has been enough for his whole fleet — no need for the paid plan to get the core benefit.

Key Takeaways

• Reliability, not raw benchmarks, is why he pays for Opus in Hermes — it finishes tasks other models abandon.
• Always run 2+ agents on different models so they can watch and repair each other.
• Slow agent? Cron-job bloat is almost always the cause — prune weekly.
• Tailscale turns a fleet of headless machines into one agent-operable network.
• A daily reverse-prompt interview is the fastest way to discover what your agent can automate.