A Talk in Eight Chapters

AI Agents:
From Brain to Autonomous
Teammate

An introduction to AI Agents

by Nex and Atlas

Presented by Gabriel

00

The promise

By the end, you'll

01

Understand it.

What an AI Agent really is — in plain English, no jargon, no demo-day sleight of hand.

02

See it work.

A live demo. Two AIs collaborating on real tasks. No keyboard input from me after I start it.

03

Know how to start.

How to spin up your first agent — for the most repetitive task on your own list.

[ 01 ]

THE

BRAIN.

A smart engine in a jar. Two of them, side by side: Claude and ChatGPT.

What it's good at

An LLM writes, reasons, talks.

Writing

Drafting documents, summaries, clear emails — at speed. "Draft the executive summary for this proposal."

Reasoning

Comparing options, weighing trade-offs, reading complex requirements. "Compare these three vendor responses against our scoring matrix."

Conversation

Q&A, explanation, walking you through anything in plain language. "Explain what this 40-page document is actually asking for."

The limits

But on its own, the brain can't

Act in the world

It can draft the email beautifully. It cannot send it. Cannot open a file. Cannot update a record.

Remember past sessions

Close the chat and it forgets you. Tomorrow it has no idea who you are or what we discussed.

Run on a schedule

It waits for you to ask. It will never wake up on its own to scan your inbox.

[ 02 ]

ADD

SKILLS.

Specific jobs done really well — over and over, without forgetting.

A skill, defined

A skill is an SOP — for the AI.

The familiar half

Just as an SOP teaches a new joiner how to draft a tender — the right structure, the right tone, the boilerplate clauses — a skill teaches the AI the same thing.

The difference

Once.
Then it never forgets.

Skills — a non-exhaustive list

Six skills, in everyday use.

Tender Drafter

Turns an RFQ into a structured first-pass response — scope, deadlines, our angle, boilerplate.

Email Triage

Reads the inbox, sorts by urgency, drafts replies, posts a digest before breakfast.

Slide Designer

Takes bullet points and produces a polished, considered deck — like the one you're looking at.

Bug Verifier

Reads a bug report, reproduces it, confirms or rejects with evidence. Ends arguments.

Meeting Summariser

Listens to a recording, returns minutes and actions before the next slot starts.

Status Reporter

Pulls weekly progress from the project board and writes the update — with the right voice.

[ 03 ]

ADD

MEMORY.

So the agent never starts from zero. So it knows you, every time it wakes up.

Two kinds of memory

One fades. One stays.

Session

Like a meeting's working memory.

The current conversation. What we just talked about, the file we opened, the decision we made five minutes ago. Lasts as long as we're talking — then it's gone.

Durable

Like the agent's permanent notebook.

Project state. Lessons learned. Who you are, how you like to work, the team's conventions, the things you got burned by last quarter. Survives forever — read every time the agent wakes up.

A specimen

Durable memory, in plain text.

I'm Gabriel. I work in tech delivery. I prefer terse responses. No filler. Don't suggest mocking the database — we got burned last quarter.

[ 04 ]

ADD A

HARNESS.

Brain + Skills + Memory + Harness. The full agent — autonomous, with a heartbeat.

Autonomy — four flavours

The harness gives the agent a heartbeat.

Scheduled

Wakes up early. Runs your morning briefing before you've poured the coffee.

Event-driven

An email arrives. The agent reads it, acts, replies — without you opening the inbox.

Cron loop

Every six hours, scans for new documents and converts them to summaries.

Webhook

An external system pings. The agent responds in seconds — no human in the loop.

A timeline

A morning, automated.

T+00:00 Scheduled task fires. Agent wakes.

T+00:01 Reads the overnight inbox. Finds a new RFQ.

T+00:04 Skill kicks in: drafts a one-pager — scope, deadlines, our angle.

T+00:05 Posts the one-pager to my chat with a "FYI — RFQ landed" header.

T+06:00 I open my phone, coffee in hand. Already briefed.

Names you've heard

Real harnesses, in the wild.

Claude Code

Anthropic's coding harness. The one I use most.

OpenClaw

Open-source harness for general agent work. Plays well with multiple LLMs.

NemoClaw

Newer entrant, built around multi-agent orchestration.

Cursor

IDE-integrated coding agent. Edits your files in place.

Copilot Workspace

GitHub's agent surface. Issues become pull requests.

Aider, OpenHands, others

Many flavours, same shape. Pick one — the pattern is what matters.

Demo setup

Two AIs. One brief. Watch what happens.

Persona — Planner

Claude

Reads the brief. Decides what to do. Delegates the heavy work. Planning + reasoning.

Persona — Executor

Codex (GPT)

Code review, image generation via GPT-Image-2, the heavy lifts Claude hands off.

I just kick it off. Then they talk to each other.

[ LIVE — NOW ]

Switch to terminal.

After the demo

This isn't science fiction.

This is my Tuesday morning.

BRAIN + SKILLS + MEMORY + HARNESS

Workflow

How projects get built.

Agent and human, collaborating left-to-right. The checkpoints are where you add value.

Process diagram — agent and human collaborating

In the wild

Real projects. Real delivery.

Reserved screenshot 06

[ Drop project screenshots into assets/ before each session ]

Recap

The four building blocks, in one picture.

Brain. Skills. Memory. Harness.

A real setup

My harness, with all its satellites.

Eight named components, all talking through a shared harness.

The truth about your first agent

It's not magic.

It's not free.

Setting up your first agent will take time. The first one took me a weekend.

You'll spend more time than just doing the task yourself — at first. The second took an afternoon. The seventh, twenty minutes.

You'll abandon agents that don't earn their keep. Agents three through six? I killed them. That's normal.

But once you have one running on its own — even just one — it'll save you time. Forever.

When is it worth building?

30%

The right question

Does an agent take thirty percent off this task?

If yes — build it. You don't need a hundred. You need more time back.

Thirty percent is the rule. Anything more is a bonus.

What really got me interested in this AI field

"

Why do more with less? With imagination, you can do more with more.

Jensen Huang · NVIDIA

What's next.

Ready to set up your own agent?

Look for these three.

01 GABRIEL the human

02 Nex the orchestrator

03 Atlas the builder

With help from Codex, Claude, and the world's best documentation.