Chapter 01 · The Basics
It sounds like a lot. It is not. One short page and the jargon loses its teeth for good. Everything after this gets easy.
01 · Start with the big ideas
Claude, GPT, Gemini. Reasons and writes brilliantly. On its own, it cannot do anything. All thought, zero reach.
A saved set of instructions the AI reuses every time, so you stop re-explaining yourself.
A model with tools, memory, and a feedback loop. One job. Follows your skills. Acts without you driving each step.
Several agents under one coordinator, each owning a piece. A small team with a lead. One goal.
The model is their raw brainpower. A skill is the one-page playbook you hand them so they work your way. The logins that let them into your systems are MCP. An agent is that hire owning one job end to end. A team of them under a lead is agentic AI.
02 · The cheat sheet
| Word | In one line | Analogy |
|---|---|---|
| Model | The thinking engine. Reasons and writes, but acts on nothing by itself. | Raw brainpower |
| Agent | A model given tools, memory, and a feedback loop to own one defined job. | A new hire, one job |
| Agentic AI | Several agents working together under one coordinator toward a goal. | A team with a lead |
| Skill | A saved instruction set the AI pulls in only when it fits the task. | A one-page playbook |
| MCP | The wiring that lets the AI reach your other tools and systems. | Keycard and logins |
| Project | A topic space with its own files and its own memory. | A labeled binder |
| Memory | Carrying things across chats. Off by default in Claude; you opt in. | Its notebook |
Do not fixate on which model. A modest one wrapped in strong scaffolding beats a brilliant one with none. The setup is the real work.
03 · How far you let it go
Each level adds power and risk together. Most people reach too far, too soon. Pick the lowest level that does the job.
Trigger fires, action runs, same every time. Zapier, Make. Example: contract signed creates a Drive folder.
Handles variation a rule cannot. Example: reads a call transcript, drafts a personalized welcome email. You send it.
"Onboard this client." It picks the steps, passes context between them. For most operators, this is the destination.
Runs end to end, pings you on exceptions only. Earn it with 30+ correct gated runs first.
Does this really need an agent? They shine on structured, repeatable work with a clear right and wrong. They wobble on anything that leans on real judgment. Knowing that line beats picking the perfect tool.
04 · Pop the hood
Someone pitching you a tool, or yours acting up? These are the parts to look for.
Talks to a model, gets usable answers, runs multi-turn reasoning.
Concrete actions: read, send, update, search.
Within a task and across sessions.
Who goes first, what is passed, what if one fails.
Defined moments a human says yes or no. Not optional.
Tools like LangChain or CrewAI add a layer between you and the AI. On a simple build that layer mostly gets in the way and makes trouble harder to trace. Reach for one only when juggling several agents leaves you no choice.
The gist
A brain. A brain with hands. A team of them. That is the whole idea.
Pick the simplest level that solves it. You can always climb later.
Does it really need an agent? If it leans on real judgment, maybe not.