WebinarNotes/raw/notes/Webinar script.md

ideas: disclamer, it's, essentially, the same demo, just with extended set of tools. your harnes is unique to you and adapdet to your own workflow

Intro

So... Let's start!

{Press start button}

---

I've always had a passion for building internal tools.

Tools that make one person's life at a company a little easier.

They cut stress here and there...

Maybe they only save five minutes a day.

No one would invest in that.

But for that one person, those could be the most stressful five minutes of their day.

Take them away, and they're happier...

...and more productive.

Back then, I'd burn whole weekends to help someone.

Myself included.

But today, things have changed...

---

We all know ChatGPT.

But what we know about it is differ from person to person.

Some says that "it just predicts the next token".

The others: "it is so capable that it will replace all jobs in nearest future"

Somehow both statements are right and wrong at the same time.

I've prepared a small journey for you.

This journey will show you how the same simple task could be done with AI as it evolved over time.

And how you can come from Chat box to your own Agentic Operating System

Mail from boss

So, let's imagine that you are working in an office and it's a beginning of your work day.

You are dreaming for the next vacation while there is a new mail from your boss.

{Open and read email}

You've got yourself a new assignment!

But, before go full in on this, you remember that you've heard about AI that can help people a lot in their work.

So you deciding to give it a shot.

Chat box

It's autumn of 2022. ChatGPT is on hype. Everyone talking about it. So it's your way to go.

You open the chatbox and typing.

"I've got this mail from my boss! what should i do? {paste mail text}"

{Chatbox gives answer}

Pretty good instructions, i would say.

AI knows about your environment and now about your task. So it can help you to compete it.

{Following AI instructions}

Task is complete. You are happy. Boss is happy.

Let's see what allowed AI to be so helpful for us.

{Open system prompt}

Each AI model has a system prompt that shapes it's behavior.

And this is exactly what we see here.

This is essentially an instruction on how to interact with our toy system.

And AI works like an advanced search engine.

It works great with

Now you just need to check on weather and do this routine till your next vacation.

AI just can't help you more. You've got the knowledge that you need. That's it.

Essentially, that is google on steroids or interactive instruction to your product, but nothing more.

Researchers of the time realized this fairly quickly.

There just should be a way to make AI to do something useful.

Allow AI to act.

ReAct

in the same autumn of 2022. ReAct technique was released.

Let's imagine that you have received this task from your boss a bit later.

And now programmers of your company integrated ReAct technique into your AI chat.

So what this means for you?

AI can act now!

let's see how it would react on the same prompt.

{use the same prompt}

AI does the job for you now!

This opened the door to some possibilities.

AI can work with information from the real world, not only from instruction.

But this technique was highly experimental and hard to use.

It required a lot of effort to add new functionality to model.

AI models were glitching, hallucinating, writing answers themselves...

But the direction was set.

This was highly usable.

let's see what allowed us to teach ai to do something useful

{show system prompt}

Tools

in summer of 2023 GPT-4 made tool definition as standard

models were trained with this definition in mind, so it became native for them.

programmers now could add functions to models the same way as they could do it in code

models could have a variety of tools for low price of adding them

so now it's even easier to integrate AI into everything

here i want to introduce a new term: Harness.

you've, probably, heard about it.

Harness is a set of tools and techniques that are built around AI model.

lot of SaaS startups working like this

just make a harness, connect to OpenAI api, take your profit

todo: here i would need some examples of saas startups

now, when we can integrate AI into more things, let's see how our work day is going

"check if there is new mail from Marcus and do what he wants"

{AI does something}

now AI has access to mail

and we don't even need to describe the task by ourselves

AI just reads it and does it

Imagine that that there is Jira ticket instead of mail

And you have task automation

"But what if AI has a question? What if the task is not clear?" - i hear you ask

{ask AI to write answer to the boss}

it can update and re-assign tickets in the same way

and the coolest thing is that, thanks to MCP servers, programmer needs less than an hour to add connection to any datasource

let's get back to our task

if we start a new session, AI would still know nothing about us.

Let's check it.

Memory

{close the terminal, open a fresh one}

"do what Marcus asked"

{AI has no idea who Marcus is}

And... it's a stranger again.

Everything we built up in that conversation — gone.

Here's the uncomfortable truth about AI models: they never remember anything.

The model's entire world is the current conversation.

Close the window — and that world is erased.

Every session, you meet the same brilliant amnesiac.

For a chatbot, that's annoying.

For a coworker, that's a deal-breaker.

You wouldn't re-onboard a new employee from zero every single morning.

So the fix had to come. And it turned out to be almost funny in its simplicity.

A notebook.

{open the Memory level}

We give AI one more tool — a notebook it can write in.

And one standing rule: if you learn something worth keeping — write it down before you answer.

Let's watch it work.

"From now on: when Kyiv is below 20 degrees, the cube goes to the top shelf. Above 20 — bottom shelf."

{AI saves a note, then does the task}

Notice — before doing the job, it quietly made a note.

Now, the moment of truth.

{close the terminal — session destroyed — open a new one}

New session. Blank conversation. Yesterday this meant total amnesia.

"Check the weather at my place"

{AI checks the weather and moves the cube}

It knows.

No explanation needed. The knowledge survived the session.

Want to see the magic trick?

{show system prompt}

There is no magic.

When a session starts, the harness simply pastes the notebook into the system prompt.

And the notebook itself?

{show memory.md on disk}

A text file. Sitting on my computer.

A text file I never wrote. AI maintains it itself.

There's a second notebook too — about me.

{show user.md — "The user lives in Kyiv"}

So I don't even have to say "Kyiv" anymore. "Check the weather at my place" is enough.

This is the moment AI stops being a tool you operate...

...and starts being a teammate.

Because a teammate remembers agreements. Remembers your preferences. Remembers you.

But look closer at that notebook.

It's tiny. On purpose.

Everything in it gets loaded into every single session — needed or not.

Facts about you fit fine.

But whole procedures? Step-by-step workflows?

Write all of those down, and the notebook becomes a phone book the AI must re-read every morning.

What we need is memory that stays on the shelf...

...and comes down only at the exact moment it's needed.

That's the next stop.

Skills

in autumn of 2025, the industry landed on an answer.

And once again, it's funny in its simplicity.

A folder with a text note in it.

They called it a skill.

Let me show you why this changes everything.

{open the Skills level — the shelf is empty}

Let's do our morning routine. But this time, I'll walk AI through it step by step.

"check the weather in Kyiv"

{AI checks the weather}

"the rule is: below 20 — top shelf, above 20 — bottom shelf. it's below right now, so move it to the top"

{AI moves the cube}

Job done. Nothing new so far.

But notice who did the thinking.

Me. I was the recipe.

The procedure lived in my head, and I dictated it, step by step.

Do this every morning — and I'm the bottleneck again.

So now, the one sentence that changes the game:

"save what we just did as a skill"

{AI writes the skill}

Let's look at what it created.

{show SKILL.md on disk}

A folder. Inside — one markdown note.

A name. A one-line description of when to use it.

And the steps we just walked through — written down as a recipe.

No code. Plain human language. My procedure, on paper.

Now watch.

{close the terminal, open a fresh one}

New session. Total amnesiac, remember?

"do the weather-based movement"

{AI loads the skill and runs it — weather, cube, done}

One line.

No briefing. No step-by-step. No me-being-the-recipe.

It found the skill on the shelf, read the recipe, and did the job.

And here's the clever part — the part the notebook couldn't do.

{show system prompt}

Look what's actually loaded: just the name and one line of description.

The full recipe stays on disk...

...until the exact moment it's needed.

That's why you can have ten skills. A hundred. Hundreds.

The shelf can be huge — the desk stays clean.

And remember: a skill is just a file.

You can read it. Fix it. Improve it.

You can send it to a colleague — and now their AI knows your procedure.

You can take it to a different AI tool tomorrow. It's plain text. It travels.

Your experience is no longer locked in your head — or in one chat window.

Memory made AI a teammate who knows you.

Skills make it a teammate with experience — one who knows how the job is done.

But one thing still bothers me.

Marcus said: "continuously — don't let it drift."

And every single run still starts the same way...

...with me. Typing.

The AI has the knowledge. It has the skill.

But I'm still the alarm clock.

What if we could remove even that?

Process

it's 2026 now. and this is where we finally arrive.

{open the Process level}

This window looks like all the others.

But it's not a chat anymore. It's a shell.

And the AI behind it is not an assistant. It's a process manager.

Watch what happens when I give it — not a task...

...but a goal.

Actually — hold on.

Before I hand over control, let me be sloppy on purpose.

"keep the cube on the right shelf"

{AI moves the cube to the right-hand shelf}

note: stage-critical — a modern model may well disambiguate "right shelf" correctly from context. Pin this beat to a deterministic response (or a low-temp on-rails prompt) so it reliably picks the right-hand shelf.

That's not what I meant.

I meant the correct shelf.

It heard: the shelf on the right.

And notice what it didn't do.

It didn't ask. It didn't hesitate. It didn't flag anything.

It just confidently did the wrong thing.

Every gap you leave, it fills.

And it fills it silently.

Right now that's harmless — I'm sitting here, I can see the cube.

But in a minute I'm going to walk away from this keyboard.

Marcus said: don't let it drift.

Turns out the first thing that drifts... is what I meant.

So let's say exactly what we mean.

"keep the cube on the right shelf: below 20 — top, above 20 — bottom. continuously."

{AI spawns a process — "started process 1"}

Look at the answer.

It didn't do the task.

It started a process. With a process id. Like a real operating system would.

And now — the most important moment of the whole journey.

I take my hands off the keyboard.

{step away; heartbeat lines tick every few seconds}

Every few seconds — a heartbeat.

Check the weather. Compare. Decide. Hold.

Nobody is typing. It just... runs.

Let's make the world change.

{drag the weather Override slider across 20°}

{the cube moves by itself}

There.

The temperature crossed the line — and the cube moved.

No prompt. No click. No me.

And it stays manageable, like any process:

"what's running?"

{ps-style table of processes}

"stop process 1"

{process killed}

Spawn. List. Kill.

Where have you heard those words before?

That's how an operating system talks about its programs.

Now — the reveal. What actually happened when I typed that goal?

{show the prompt the shell wrote for the worker}

The AI wrote... a prompt.

For another AI.

It authored the worker's instructions — like a manager writing a job description — and launched it.

AI managing AI. And you? You just state the goal.

Remember what Marcus asked for? "Continuously. Don't let it drift."

Handled.

I haven't touched the keyboard in two minutes.

You can finally get back to dreaming about that vacation.

But... one thing still feels unfinished.

Look at this window. It's still a terminal.

I still had to type the goal. To talk to it the way a programmer talks to a shell.

Most people never will.

What if this whole routine could stop being a conversation at all...

...and become a tool? A small one. Made for exactly one person.

OS

{open the OS app}

Look at this window.

No chat box. No blinking cursor. Nothing to type.

A status bar. One button: "Check temperature in Kyiv". One checkbox: "Do every 5 seconds".

That's the whole interface.

{click the button}

{AI checks the weather, the cube slides to the right shelf, the status bar shows its one-line report}

One click.

The agent checked the weather, applied Marcus's rule, placed the cube — and reported back in one line.

But wait. Where did the rule go? I never typed it.

It's baked in. This app was built around Marcus's instruction.

The prompt was written once — and disappeared behind a button.

{tick "Do every 5 seconds"}

And now it's not even a button anymore.

{drag the Override slider across 20° — the cube crosses on its own; the status bar updates}

It's an appliance. It just... works.

Notice what disappeared along the way.

The conversation.

There's still a full AI agent in there — same model, same tools, reasoning on every tick.

But you don't chat with it anymore.

You click it. You tick it. You close it.

You interact with it the way you interact with any other program on your computer.

The agent became... a program.

A tiny program that does exactly one job. For exactly one person. You.

Remember where we started tonight?

I told you I used to burn whole weekends building little tools like this.

Tools that save one person five stressful minutes a day.

This one took 10 minutes.

And I didn't write it — I asked for it.

That's the last step of the ladder.

The chat box didn't just get smarter.

It dissolved — into the operating system.

Into little tools you make for yourself.

---

And here's the secret of tonight's whole journey.

The model never changed.

Chat box, ReAct, tools, memory, skills, processes — even that button — behind every level, the same AI model.

What changed was everything around it.

The tools it can reach. The notebook it keeps. The skills on its shelf. The processes it runs. The buttons it hides behind.

That's the harness. And the harness is the whole difference...

...between a stranger in a chat box and an operating system that works while you don't.

And here's the part that matters for you:

nobody can sell you this off the shelf.

Because the harness is built from your mail, your rules, your procedures, your routine.

It grows out of the way you already work.

Your harness is unique to you.

You don't buy it. You build it — one small tool at a time.

---

And don't take my word for why this matters now.

Stanford measured it.

Their researchers tracked 46 teams working with AI — matched against 46 similar teams without it — for more than two years.

The teams that really learned it pulled away from the ones that just... had it.

In spring 2023, the spread between them was under five percent.

By summer 2025 — nineteen.

The gap quadrupled in two years. And the curve is still bending upward.

And remember — everyone had the same models the whole time.

The difference was never the model.

It was who built something around it.

note: source — Stanford SWEPR, difference-in-differences analysis, Apr 2023 → Jul 2025 (see wiki/sources/2026-07-30-stanford-swepr-widening-gap.md). Honest caveats if asked in Q&A: talk-published, not yet peer-reviewed; measures software teams, not office workers; Stanford says "quality of usage" decides, without naming which practice.

---

We started this journey by pasting an email into a chat box.

We're ending it with a button that already knows what the email said.

From a chat box... to your own Agentic Operating System.

Notes

• push yourself up the management chain
• terminal, MS office is skeuomorphism

Andrej Karpathy

This is a new paradigm for interacting with Claude that is significantly more "inline" with all the other human activity org-wide. Once you do all of the under the hood engineering work to make this "just work" (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads. Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It really takes a while to wrap your head around it, but it works and it is awesome.