100 hours of Hermes Agent lessons in 19 minutes
Nine practical lessons from 100 hours with Hermes Agent: model choice, failover agents, multi-device setup, performance tuning, and daily workflows to get more done wit
Hermes Agent can function as a full-time AI employee when used correctly. Used incorrectly, it performs like a standard chatbot. The following nine lessons are drawn from hundreds of hours of use over the past four months.
1. Choose the Right Model
Opus is the strongest model available for Hermes Agent. In agentic tasks, Opus completes assigned work reliably. Other models are more likely to stop partway through a task and fail to finish it.
Opus is expensive. The author spent 40 per day. Spending varies. Some days involve no spending at all. The author is willing to double current spending on Opus because Hermes runs the business.
Lower-cost options exist:
- ChatGPT 5.5 and later: Suitable for users who already pay for a ChatGPT subscription. It is not at Opus level, makes errors, and is less capable at tasks. Versions before 5.5 were not usable.
- GLM 5.2: The lowest-cost option, at a fraction of the price of the other two. It is more robotic in tone, likely because it is distilled from these models, but it completes tasks and is very inexpensive.
Recommendation: Use Opus for the highest reliability. Use ChatGPT if a subscription already exists. Use GLM 5.2 to minimize costs. Businesses with meaningful revenue should prioritize Opus, since guaranteed task completion justifies the cost.
2. Run at Least Two Hermes Agents
Running a single agent is insufficient. Agents can fail, and a backup allows recovery.
The author runs a main agent named Hermes on Opus and a second agent named GPTme's on ChatGPT. Hermes breaks less often than OpenClaw, but both can fail. When the ChatGPT account connected to GPTme's went down because a token expired, the main agent repaired it. The author sent a screenshot and the error description to Hermes, which fixed the problem automatically.
Agents can monitor and repair one another. Users should connect additional accounts, such as ChatGPT, Gemini, or Google, to create more agents. OpenClaw can serve as an additional backup and offers features that other agents may lack.
Creating a new agent (profile): In Hermes, agents are called profiles. Two methods are available:
- Through an existing agent: Instruct the agent to create a new Hermes profile, specify the model, and optionally assign a name. For example, "Name it Harry."
- Through the Hermes dashboard: Open the dashboard, select Profiles, click Create in the top right corner, choose a model, enter a name, and launch the profile.
3. Do Not Create Separate Hardware or Accounts for Each Agent
Some users buy a dedicated computer for each agent and create separate Gmail, Apple, or other accounts for it. The author considers this unnecessary. It adds friction without a meaningful gain in security.
Hermes Agent performs only the tasks it is given. It is not sentient, and it does not act outside its instructions. The author knows of no one who has experienced a security incident with Hermes. Most security concerns are overstated.
Recommendation: Run agents on the main computer and primary accounts. Exercise personal accountability and consider carefully what each instruction asks the agent to do.
4. Use the Right Platform for Each Task
The author uses three platforms:
- Hermes Desktop for work at a computer. It offers the best user experience for an AI agent. Users can switch between profiles, view all cron jobs, and manage multiple pinned sessions at once.
- Telegram for deep, multi-threaded work on a phone.
- iMessage for quick tasks on the go. Hermes supports iMessage out of the box in its latest update. The author saved a contact named Hermes and pinned it to the top of the message list.
5. Use Telegram's Formatting for Tables
Telegram now supports tables, bold text, paragraphs, and other formatting in messages. Users can schedule cron jobs that return structured results. For example, the author runs a daily stock research task on leading AI stocks and receives the results as a table with ratings and supporting information each morning. Similar tables can display top-performing content or priority emails.
6. Remove Unused Cron Jobs to Improve Performance
Slow or unresponsive agents are often caused by cron jobs. Cron jobs are scheduled tasks that run in the background. They are easy to create and difficult to track, and many users forget the ones they have scheduled.
To fix performance problems:
- Open the Hermes Desktop app.
- Select the cron button to view the list of all cron jobs.
- Pause any job that is no longer needed.
- Repeat this review regularly, such as once a week.
Removing unused jobs improves agent speed and reduces token usage. Many cron jobs consume more tokens than expected.
7. Use Tailscale to Connect All Devices
Tailscale is free and creates a private network across all devices. Users with more than one device, such as a computer and a phone, Mac minis, or DGX Sparks, should install it. The author has used the free version continuously. A paid version exists, but its features are not covered here.
With Tailscale, Hermes Agent can operate across all connected devices. Examples include:
- Asking Hermes to list the local AI models running on a DGX Spark. Hermes connects to that device over SSH, checks the running models, and reports the results.
- Asking Hermes to download and run a new local model on another device.
- Asking Hermes to retrieve a presentation from another computer and send it to the user.
- Accessing a localhost application on the main computer from a phone, which allows testing on the go.
The author runs five computers without monitors. The Mac Studio is the only device connected to a display, and Hermes operates the other machines remotely.
8. Use Reverse Prompting to Identify Useful Tasks
Reverse prompting instructs the agent to ask the user questions. The answers allow the agent to identify tasks it can perform.
The author conducts a morning interview with Hermes. The agent asks about daily priorities, current tasks, sources of stress, and existing workload. It then determines which tasks it can take over, which it can automate, and which require the user's permission. The interview takes about five minutes and saves considerable time.
The author recommends running the interview every morning. Each session typically reveals two to three new tasks suitable for the agent.
9. Assign Tasks Through the Kanban Board
Tasks identified during the morning interview can be moved to a Kanban board, where the agent completes them.
To open the board:
- Open a terminal. Ghostty or the built-in terminal can be used.
- Type
Hermes-dashboardand press Enter. A website opens. - Select Kanban board in the bottom left corner.
On the board, users can add tasks, move them between columns, and assign them to the agent. The agent then begins working on assigned tasks. The board provides a single view of tasks for both the user and the agent.
