Agents 101 for Normies — Interactive Demo

Agent 101 for Normies

What AI agents are, how to set one up, and what's happening behind the scenes.

Who are you?

A normal person not working in the tech industry — or in a non-software role at a tech company.
Not a vibe-coder (yet). Despite the trashy name, this is a useful activity — let me make the pitch.
You feel like a postal worker in 1985 learning about email for the first time.

What this is (and isn’t)

This is not another vibe-coding course. The goal is to become a knowledgeable, discerning buyer of the ever-growing stack of “AI” tools — not a developer of them.

We do that by:

Building small applications with AI that improve the efficiency of our day-to-day work
Getting a working understanding of what makes AI systems tick — model capabilities, limitations, memory, eval, guardrails
Learning by constructing, not consuming — I have no interest in rebuilding what exists. I want to spend my money wisely, buy tools I understand deeply, accomplish my goals, and turn them off.

What will we cover?

A use case you can set up in 15–30 minutes, what it reveals about how agents work, and a roadmap to go deeper.

Start
My Day

Build
a Skill

Modes
of Use

Under the
Hood

Jargon
Guide

★

Context
Rot

What’s out of scope

Building AI for other people (apps, products, platforms)
Building custom tools (custom MCPs, harness development)
Agent-as-product (custom agents, A2A protocols, etc.)

Start My Day

One command. Six tools. Your morning briefing — done.

It's a Skill that an “Agent” applies. At the most fundamental level:

Skill = SOP + Trigger + AI

SOP

How you'd teach a new hire to do a specific task, written in plain English with numbered steps and documented edge cases.

Trigger

So the new hire (or AI) knows when to follow the procedure. A command, a schedule, or an event.

The AI model knows how to execute the SOP and use the tools you specified.

Up next: We’re switching to a live Claude session to build this skill from scratch. Follow along.

Let's Build the Skill

You need a clear SOP before you automate anything, with humans or AI.

Your tools

Slack, Gmail, Google Calendar, Jira, Zendesk, My News App.
Every day, for the first 30 minutes, do the following:

💬

Sweep Slack

DMs, channels, threads, group chats. Flag anything needing action.

📧

Sweep Email

Priority inbox. Identify actionable items.

🎫

Create Tickets

Collaboration items become Jira tickets. High-confidence: auto-create. Low: ask first.

📋

Check Jira

Read colleague comments. Identify actions needed from me.

🎧

Check Zendesk

Complaints, requests, and defects assigned to me.

📰

Fetch News

Domain and industry digest.

Set up your Skill in your AI agent tool

Install plugins and connectors for each tool. Most software you use at work already has plugins (or will within 6 months). If not, your IT team can likely build a custom connector.
Write the prompt: tell the AI the trigger (/start-my-day), the goal (summarize what I need to do today), success criteria (all action items identified), tools, and your SOP above.
Test it. Review the output. Refine the SOP. Repeat.

This pattern works everywhere. Claude calls it a Skill. ChatGPT calls it a Custom GPT. Gemini calls it a Gem. The name doesn’t matter. The pattern does: write clear instructions with a goal, steps, tools, and what “done” looks like.

Prompt Template

Build me a skill called "start-my-day".

Trigger: /start-my-day
Goal: Summarize everything I need to act on today.
Success criteria: Every actionable item across all tools is surfaced. I'd rather see a false positive than miss something.

Tools available (already installed):
- Slack
- Gmail
- Google Calendar
- Jira
- Zendesk
- My news digest app

SOP — run these steps in order:

1. Sweep Slack — DMs, channels, threads, group chats. Identify anything that requires action from me.
2. Sweep emails — focus on priority inbox. Flag anything with an actionable item.
3. Create tickets — items requiring collaboration become Jira tickets. High-confidence items: auto-create. Low-confidence: ask me individually.
4. Check Jira — read comments from colleagues, identify actions I need to take.
5. Check Zendesk — complaints, requests, and defects related to me.
6. Fetch news digest — read up on my domain and industry.

Output format: group by urgency (act now / act today / FYI). Include source and link for every item.

Before you start building, ask me clarifying questions about anything that's unclear or underspecified.

Under the Hood

First, how the app wraps the AI. Then, how agents loop.

Phase 1: The Harness Phase 2: The Agent

Your Computer

AI Company's Server (Anthropic / OpenAI)

You Type a message

THE APP (Claude Desktop / ChatGPT)

Your Files Project files
Uploaded docs
Local folders

Chat Window Text box, buttons, history

Message Packager Bundles conversation + your files

Internet Connection Sends everything to the AI company (API call)

Show the Answer

The AI Brain Reads your conversation and files.
Picks the best next words.
Sends back one response.

Has no memory between calls.
Sees only what the app sends.

⚠

No memory box.
No knowledge base.
Every call starts from zero.

Notice what's missing.

There is no memory box. No knowledge base. The AI starts from zero every single call. The app fakes continuity by re-sending everything — your conversation and your files, every time.

Use the buttons below to walk through each step.

Modes of Use

How much you set up vs. how much AI does on its own

Setup investment vs. AI autonomy by mode

Your effort per task

AI autonomy (how much it does on its own)

Click a mode to explore

When to use

Try this

What you need

Keeping AI focused (context hygiene)

Mode 0

One-off

a search engine that talks back

Mode 1

Chat

a brainstorm partner

Mode 2

Project

a colleague you've already briefed

Mode 3

Agent

an assistant running your playbook

Mode 4

Orchestration

a system that runs without you

The 80/20 trap. AI makes the first 80% look effortless — and it genuinely is. But the last 20% (reliability, edge cases, integration, maintenance) is where the expertise lives. AI didn’t eliminate that work. It made it invisible until you’re already committed.

Beyond this session

Building AI Products

RAG · memory · guardrails · agent orchestration

Building AI Models

eval · fine-tuning · training data · RLHF

Jargon Guide

Buzzwords you’ll hear, what they actually mean, and whether you need to care.

🔌

MCP (Model Context Protocol)

A standard way to connect AI to your tools. Like a USB port: one plug shape that works with Slack, Gmail, Jira, and everything else. Without it, each tool needs its own custom integration.

You don’t build MCPs. You install them like browser extensions.

🔄

Agent

An AI that takes action, not just answers. It reads your instructions (the Skill), uses tools, checks results, and keeps going until the job is done. A chatbot is one turn. An agent is a loop.

You’ve been using one this whole talk.

📜

Skill (SOP + Trigger)

A written procedure with a trigger phrase — step-by-step instructions an AI agent follows to complete a task. You define the goal, the steps, and the tools. The agent executes. “Skill” is the reusable playbook; “agent” is the system that runs it. Same pattern, different names: ChatGPT → Custom GPT, Gemini → Gem.

You built one in the “Build a Skill” panel. The pattern transfers to any product.

📦

App (Wrapper / Harness)

The app that wraps the AI model: ChatGPT, Claude Desktop, Copilot, etc. It manages your chat window, files, and tool connections, then sends everything to the AI. The AI itself has no memory — the app fakes it by re-sending your full conversation every time.

The “How AI Works” diagram shows exactly what it does behind the scenes.

🗂️

Context (Context Window)

Everything the AI can see in a single call — your conversation history, uploaded files, and instructions, bundled together. The AI has no memory; context is all it gets. When it fills up, older details get pushed out and quality drops.

The bonus panel shows exactly how this breaks down.

📡

API (Application Programming Interface)

How software talks to other software over the internet. When you send a message, the app makes an “API call” — it packages your conversation and sends it to the AI company’s server. You never see this; the app handles it.

Step 5 in “How AI Works” — the arrow leaving your computer.

🔗

Orchestration

Multiple agents working together, each on a different piece. One does research, another writes, a third reviews. Like a team of specialists with a project manager.

Out of scope for beginners. Know it exists; don’t worry about it yet.

📚

RAG (Retrieval-Augmented Generation)

The AI searches your documents first, then answers based on what it found, instead of guessing from memory. Like handing someone a reference binder before asking them a question.

Useful when you hear “it can search your docs.” That’s RAG.

📏

Tokens

How AI measures text. Roughly 1 token ≈ ¾ of a word. Every AI model has a token limit, the maximum it can read and write in one go. Hit the limit and older context gets dropped.

You saw this in the Context Rot demo. It’s why long chats degrade.

🎯

Fine-tuning

Training an AI model on your own data so it gets permanently better at a specific task. Expensive, slow, and rarely needed. Different from giving it instructions (that's prompting).

You almost certainly don’t need this. Prompting + tools gets you 95% there.

✏️

Prompt Engineering

Writing clear instructions for AI. Same skill as writing a good email to a new hire. Be specific about what you want, in what format, and what “done” looks like.

You just did this when you wrote your Skill prompt.

🛒

AI Products (ChatGPT, Claude, Gemini, Copilot)

What they share: a chat interface, API access to models, and project workspaces. Key differences: Claude has Skills/MCPs + Code (terminal agent) + CoWork (teams). ChatGPT has desktop agents + GPTs + custom actions. Gemini has deep Google integration.

Pick one, learn it deeply. They all access similar underlying models.

🗃️

Claude Flavors (Chat, Code, CoWork)

Chat = web interface (claude.ai), good for Modes 0–2. Code = terminal/IDE agent, good for Modes 2–4. CoWork = team workspace with shared projects and knowledge.

Start with Chat. Move to Code when you’re ready for agents.