⚖️ With great power comes great responsibility
In Tracks 1 and 2, you saw Hermes gain capabilities: computer actions, integrations, MCPs, memory. Each of these is a door to the real world. A chatbot that makes a mistake gives you a bad answer; an agent that makes a mistake can send the wrong email to the wrong client or delete a file. The difference isn’t one of degree — it’s one of kind.
🧭 The right mindset
Think of the agent as a very fast, very literal intern. You wouldn't give a first-day intern the company's bank password. You give them small tasks, observe them, and only expand their access when you trust them.
- •Enable capabilities one at a time, not all of them during setup.
- •Observe its behavior before giving it access to irreversible actions.
- •Treat each new capability as a conscious decision, not a default.
Errors have real-world consequences.
One capability at a time.
Trust it after you see it in action.
Every capability is a decision.
🔐 The principle of least access
In security, this has a name: least privilege (least privilege). The rule is simple: each part of a system should have the minimum access necessary to do its job — and nothing more. This applies to Hermes, your phone, and your bank account. It’s the most universal security principle there is.
Illustrative diagram · each ring outward means more power and more risk
📐 The three access levels
- Read — the agent only observes (check email, read a file). Low risk, reversible.
- Write / edit — creates or modifies (draft an email, save a file). Medium risk.
- Send / delete — irreversible actions in the world. High risk; save for last.
📧 The Email-Sending Case
The classic example, and where most people got burned: you connect Gmail and, excited, grant everything — including sending. In an automatic morning briefing, the agent misinterprets an instruction and sends an email to the wrong list. You can’t “unsend” an email.
✓ Start here
- ✓Connect read-only of email (summarize inbox).
- ✓Let the agent draft answers for you to review.
- ✓Observe for days to see if the readings are correct.
✗ Avoid at the start
- ✗Allow automatic sending without human review.
- ✗Grant send access within an unsupervised 24/7 cron job.
- ✗Trusting "it understood" without testing it in read-only mode.
💡 Practical tip
Golden rule: irreversible actions last. Email, transfer, deletion — only after the agent has proven, in read/draft mode, that it understands what you want.
🔑 Never paste API keys into chat
Remember memory (Trail 2)? Every message from every session is indexed and savedIf you paste an API key into chat, it becomes a permanent part of memory — even worse if you enabled the daily backup on GitHub: now the secret is versioned. A secret belongs in a environment variable, not the chat.
The right way to add a key (illustrative)
🚨 Caution
If you've already pasted a key into the chat, consider it compromised: rotate it immediately. When you rotate it, the old key stops working forever — exactly what you want.
🚦 Allow once / session / never
When the agent needs a new permission, it asks. You choose between three levels — and each trades convenience for control. Think of them as a gate with three openings.
Allow once — allow one time
For something unfamiliar or one-off
The agent performs the action this time and asks again next time. Maximum control, minimum convenience. Use it for things you don't trust it with yet.
Allow session — allow for the session
For repetitive tasks in the same job
It lasts for the session; after a reset, it asks again. A good balance for a task that repeats the same action several times.
Never — block it completely
For what's dangerous or unwanted
The agent never asks or executes again. Use this for actions you've decided it should never take.
✅ Responsible use applies to all AI
The best thing about security is that it’s transferableThe habits you build with Hermes — least access, secrets kept out of chat, irreversible actions saved for last — protect you with any agent, plugin, or AI system you use from now on.
🧠 Practical tip
Before enabling any new capability, ask three questions: (1) is it reversible? (2) what’s the worst-case scenario? (3) Have I tested it in read-only mode? If anything gets stuck, it’s not ready yet.
Only what’s necessary.
Env, never chat.
Send first, delete later.
Applies to all AI.
📌 Module Summary
Next Module:
3.2 - ⭐ Northstar (Goals): give the agent a direction to pursue until it reaches the goal.