2 / 13My First AI AgentPart I · 25 min read

The AI Cheat Sheet

Speak the language of agents. First pass = the short “remember” lines; details wait for later lessons.

How to read this lesson

This is the dictionary for the course. Read the intro and the remember line on each card, then move on.

  • Come back to a card when a later lesson names that word (install, SOUL, tools…).
You do not need to memorize every term today. You need four pillars (stack), the loop, and “chat vs harness.” The rest is there so Settings and docs stop sounding like alien code.

1 · Core stack

Four words name almost everything in this course.

LLM

Large Language Model — predicts the next bit of text. Pure chat is mostly text in → text out. It can sound smart without ever touching your files. In an agent setup it is only the brain.

LLM = brain only.

Agent

A system: LLM brain plus tools and memory, driven by a loop the app runs (Think → Act → Observe). The loop is around the model, not a secret chip inside the weights.

Agent = brain + tools + loop.

Harness

The app frame that runs the loop: messages, tools, approvals, memory, gateway. Hermes Desktop is that frame on your machine — what you control.

Harness = the app that runs the loop.

Runtime

Where the process lives: spare PC, VPS, container… We prefer a dedicated machine, not the laptop that holds your whole life. Different idea from “brain” or “app design.”

Runtime = the machine.

One line: LLM thinks · harness runs tools · agent is the whole loop · runtime is the box. Next: the loop itself.

2 · The agent loop (not the LLM alone)

The loop is not “inside the LLM alone.” Operators decide who does Think vs Act vs Observe.

Three beats. Think (model plans) → Act (app runs a tool or answers) → Observe (result goes back into the chat for the next Think).

That cycle is how an agent system works. It is not a hidden loop living only inside the LLM. Chat-in-a-browser without real host tools is mostly Think → reply.

Agent loop — who does each step
01
Think
LLM plans
02
Act
Harness runs tool
03
Observe
Harness → messages
04
Again
LLM thinks next

Think

The model plans: what next, which tool, when to stop. You usually see this as the next reply or a tool request — not a separate “loop chip” inside the model.

Think = model plans.

Act

Something happens outside pure text: write a file, run a command, call a service — or just answer. Approvals and host safety apply here.

Act = app (or final answer).

Observe

The tool result or error is pasted back into the chat so the next Think is grounded. No observe → higher chance of invention.

Observe = result back into chat.

Planning

Choosing steps, then re-planning after each Observe. You help with a clear goal, constraints, and “done when.” (Fine print; the three steps above are enough for pass one.)

Plan → act → check → re-plan.

WhoBeatWhat happens
Think · LLMPlans the next moveMaybe asks for a tool, or drafts a final answer. Think lives in the model’s reply.
Act · HarnessActually runs the tool — or sends the final answerHarness (+ tools on the machine). Approvals may stop a risky act here.
Observe · Harness → LLMPuts the tool result (or error) back into the conversationThat observation feeds the next Think. Errors count as observations too.
Think = model plans. Act = app (or final answer). Observe = result back into chat.
Mnemonic: LLM thinks · harness acts · harness observes · LLM thinks again. That whole cycle is the agent.
  • I can walk Think → Act → Observe once out loud
  • I assign Think to the LLM and Act/Observe packaging to the harness
Exercise · Check yourself

Who actually runs a file-write tool call?

Pick one — the answer is revealed straight away.

Exercise · Check yourself

A tool comes back with “permission denied.” What is that?

Pick one — the answer is revealed straight away.

Sources

3 · Models, cost, and keys

How you pay for the brain, and where secrets live. One cloud rule: conversation content can leave your machine.

Token

Billing unit for model input/output — roughly part of a word. Free tiers still have limits.

Tokens ≈ cost meter.

Cloud model

Brain runs at a provider instead of on your machine. Convenient; prompts and replies leave the host. You pick a provider during install (lesson 03).

Cloud = data can leave.

Local model

Weights run on hardware you control. Different cost (GPU/time). Optional later — not day one.

Local = stays on your side.

API key

Secret that proves “this is me” to a provider. If it leaks, someone can spend in your name.

Keys never in chat or git.

`.env`

Local file for secrets, kept out of git. Programs read keys from there.

Secrets in .env, not notes.

4 · What it remembers, and what it forgets

Close the chat and today’s conversation is gone. Anything the agent should still know next week has to be written to a file. That one difference — context vs memory — is what beginners get wrong most, and it explains half the “why did it forget?” moments ahead.

Context

What is in the current chat window for the model — session “RAM.” Close or compress the thread and it fades.

Context = this session.

Memory

Facts you (or the agent) write to disk so they can load again next week — e.g. MEMORY.md. Different from “it was in the chat yesterday.”

Memory = written down.

SOUL

Plain idea: a short identity file — name, tone, hard limits (“never delete without asking”). Lives with the Hermes profile on the host. You will set this in lesson 04; for now just know the word means “who this agent is supposed to be.”

SOUL = personality file (lesson 04).

Hallucination

When evidence is missing, the model still sounds sure. Files and tool results reduce that risk — they don’t delete it.

No evidence → invention risk.

Context vs memory is the mix-up that bites beginners most. SOUL is just “standing personality” — full how-to comes later.

5 · Tools, skills, plugins, and MCP

Two words carry most of it: tool = something the app can do (write a file, list a folder). Skill = a howto package that teaches the agent how to use tools for a job. Everything else can wait.

WordWhat it meansMeet later when…
ToolCallable action — the Act stepYou open Tools in Desktop (lesson 06)
SkillHowto / procedure file (SKILL.md)You trim or add skills (lesson 09)
PluginExtra UI inside Hermes DesktopYou enable Desktop extras
MCPPlug for an outside service (docs, repos…)You need a live external system — not day one

Tool

Something Hermes can run: a name, some inputs, a result or error. Hands of the agent, listed in the Hermes tools docs. Output feeds Observe.

Tool = do it.

Tool call

The model asks for a tool with structured inputs; Hermes runs it (sometimes after you approve); the result comes back as a tool message.

Ask → run → result → think again.

Skill

A reusable howto (often SKILL.md). Helps Think pick good steps. It does not replace tools — it teaches when and how to use them.

Skill = howto.

Plugin

Optional Desktop add-on (extra UI or product feature). Not required to finish Part I.

Plugin = Desktop extra (later).

MCP

A standard way to plug external services. Powerful and another trust surface. You can finish early lessons without MCP.

MCP = external plug (later).

If install “everything,” empty sessions get expensive and the model thrashes. Prefer few tools, few skills, clear jobs.

6 · Reach outside the Desktop window

Two words for work that happens when you are not staring at the app. Full how-to is lessons 05 and 11.

Gateway

Bridge so the harness can answer from a messaging app (Telegram and others). Only allowlisted people may talk. Setup is lesson 05.

Gateway = pocket access (later).

Cron

A job on a schedule (e.g. weekday morning summary). Needs a host that is awake. Lesson 11.

Cron = scheduled job (later).

Chat product vs harness agent

Modern browser chat (ChatGPT and friends) can already use some tools. The useful contrast is not “they never touch tools” — it is who owns the loop, the files, and the host.

Browser chat (e.g. ChatGPT)Agent on your harness
Tools are their product features, on their serversTools run under your harness on your runtime (files, shell, …)
History lives in their account — you export if they let youFacts and receipts can live as files you open offline
Limits and policies are vendor defaultsIdentity, approvals, allowlists are yours to set (SOUL, security)
Great for quick answers and hosted workflowsBuilt for persistence, multi-step work, and local control
Motivation, not trash-talk: when you need your disk, your rules, and a loop you can audit, you want an agent on a harness you own — not only a tab that forgets when the account does.
From chat to agent
Same kind of brain · different ownership
Browser chat
Vendor tools & history
On their servers
Fast answers
Less local control
Harness on your host
Your loop · your tools
Think → Act → Observe
Files · rules · receipts
Open offline
Checklist · this device0/5

You are done when

Proof · this device0/1

Stored on this device only — not sent anywhere.