Introduction

I think I talked, well, many of you weren't here last month, so I introduced this concept of a personal operating system.

From Personal to Company Operating Systems

And actually, that's what Melisandre was just showcasing. We've got this thing called the personal operating system and a company operating system. The idea is like more and more as people, as we start to get enhanced by agents, you need to have some kind of structure that allows you to help the individual get work done.

but the individual is in the context of a company and a whole bunch of other people that each have their respective agents that need to also execute and do the things that they're doing and

It's a little bit disconcerting that I can't have my screen doubled because I think that was what was messing it up before

Behind the Scenes: Files, Prompts, and Structure

So what we have here is Is one of those massive prompts, but the starting point actually let me see if I can collapse all of this first first.

This is what the top level looks like.

A Simple File-Based Setup

What you're seeing at the top left there is just a file structure, by the way. It's literally just folders and files on my machine.

Text files. Nothing else.

And basically those text files are the instructions that you give the AI.

Core Instruction Files and Context

And the first instruction is this bit here, which is actually in my chief of staff It has a whole bunch of context on what my operating system looks like. I also have a little bit of trouble trying to read it.

It starts with a little bit of context of you've got your chief of staff and companies and then you've got the operating system here. Each one of those has different entities.

In my case, if I go into companies, you'll see Mindstone. and then we have, this is actually an old version of the operating system,

we have the exec team and we've got everyone else, and so that's just permission -based.

These are actually folders on a Google Drive, for what it's worth, with different people having access to different folders. That's, yeah, all it is.

Tooling via the Model Context Protocol (MCP)

so for those of you that aren't familiar with MCP, Model Context Protocol, it's basically a way to give your AI access to different tools.

You then have a description of how it can use tools, description that I'm giving to the AI, how do you use tools, what are the types types of tools that you have available and so on.

Tools being read my email, draft an email, send an email, look at my calendar, update a calendar event, like all of those are different tools that you can give an AI there they go through a protocol called MCP and here is a little bit of a

Sub-agents and Memory Architecture

You then have sub -agents which allow the agent to execute other agents so that they can go and orchestrate and do bigger tasks.

It's got memory, so everything in relation to, actually let's see here, so you've got in memory here, you actually have all of the different topics and you've got core.

before, this mimics short -term memory and long -term memory, and the idea is that at every prompt, every instruction I give the AI, it reads in the short -term memory, which is basically a file on me, with some of the projects that are outstanding, and I'm not

going to leave that too long because I have no idea what the hell you're going to see in there at the moment.

And then long -term memory, it decides, the AI decides, is there something else that I should be looking at for whatever task that it has, right? So it will scan my memory and figure out, are there other bits that it needs to execute on?

Skills and Playbooks Across the Organization

Now, where things start to come together is in skills. And this is part of what Melisande was showing before, is that the AI has got different skills. My chief of staff has all of these skills that I wrote.

So now in MindStone, we then have the same thing here. Well, this is actually a new one. So if I go, because I haven't set up the latest version of ours at the moment.

Where do we have? We used to call them playbooks. So that's where they are then. And they're then by team.

Team Playbooks (now Skills) and Sharing Know-How

This means that we can now share different playbooks, which means that if I wanted to go and write a marketing or add a marketing collateral somewhere, It will go and find the right playbook that was written by the person in marketing who has clearly outlined how we produce marketing collateral, for example.

Developer View vs. User Experience

Okay, now, a lot of this is fairly technical.

This looks like Cursor, right? I mean, it's actually fairly similar.

You've got the text in the middle. On the right -hand side, you have the AI that would execute.

A Friendlier Interface on Top of the Agent Stack

But the thing that we recently did was this. And this is the thing and this is literally the very first time I'm showing this by the way, so hopefully it will work Basically what we did is we tried to build an interface on top of this whole thing

That it could just execute for people that are less technical and that is less scary, but ultimately if I look here You get the same thing. You're just seeing the folder structure

It will read in the agents .md, which is the exact same file that I just just showed you, and when I start putting an instruction in, it will just execute on that instruction.

Demo: Meeting Prep Workflow

So here, let's see, what have I gone through? Let's go through meeting prep for all my meetings for tomorrow.

You can see the interface here is fairly similar, or at least we're trying to make it as easy as possible. Available.

So here I long pressed, which then launches it into voice mode. So first thing it does, it figures out...

Tool Discovery and Calendar Context

Now let me get the list of tools available and search for meeting prep skills. Let me use the bash tool to list the skills directory and then get your calendar for tomorrow. Let me read the external meeting prep skill and get your calendar more efficiently. Let me discover the right category for calendar access.

So... let me get tomorrow's meetings using the meeting notes manager and discover available calendar actions i wonder if it's continuing when i go into i can see from the meeting notes manager that there are no meetings tomorrow november 19th 2025 i'm going to quit that here you can see

in more detail what is happening with the actual agent so you can see here at step seven and it read six files, so you can see the files that it actually read, what they related to. Step eight, it read another nine files.

What are we looking at? So it seems to be going through meeting notes, so meeting notes are stored in my system ahead of time.

Okay, so it's basically now it's just going through the execution. It's not perfect yet, because like this, Ah, wait, there we go, we got something.

Reading Meeting Notes and Launching Research

So these are the meetings that are currently there that need prep, and now let's see what it's gonna do. Read the meeting files to see which ones need prep, and then here, okay, now it should go

and launch deep research on basically all the attendants. But what it's now going to do is launch a sub -agent that is going to do research on every individual that I'm meeting, the companies that they represent,

as well as look at my emails in order to figure out what was the purpose of the meeting, which then helps it tweak the prep that goes through.

Storing Prep Outputs

And then at the end, it will actually take that and it will store it in my meeting prep, in my meeting tool, which, interesting enough, looks like this. And then you end up with prep notes there.

So it will go and populate this, for example, which is a meeting for tomorrow.

Parallelizing Tasks

So the other thing here is you can just run them in parallel. So as that is working, let's try another one.

Live Demo: Using Limitless Logs and Phone Automation

I've been using, for those of you that were here, again, last month will have seen this, but I'm using this thing called Limitless. I don't know if anyone here has ever seen that,

but it's basically it records what you're saying at any point, and then my AI agent has access to that at the same time. And so, let's see here.

Instruction: Analyze Logs and Call with a Prompt

Can you go through my limitless logs of the last hour based on the talks that happened at the AI meetup tonight? And then call me and ask me a question I can ask the audience here that would help them kind of think through what is happening with AI in relation to what was demoed today. day.

Now I'm wondering if it has my phone number, but we'll see.

So it's going through the same process.

Running Multiple Agents at Once

Now you can see that the agents are running in parallel. So I'm doing this whilst it's actually preparing my meetings for tomorrow as we stand, and I can actually

go here and see where that is at. So here it's actually, you can see I'll launch these

is a subagent, so this is going to take probably 10 minutes or so for it to go and launch these subagents in parallel to go and treat each one of them individually.

Nomenclature Gotcha: Playbooks → Skills

And then here, I wanted to add one more thing, which is make sure to use your phone automation skill for this.

Pla a skill. Great.

This is because we changed nominations from playbook to skills and that was literally a week and a half ago, so in my head I still call them playbooks.

And maybe I add one more thing. I might actually type this.

Don't go and spam me please.

Refining Scope to the Correct Time Window

Okay, so it found, interesting, so it found a demo from earlier today because I did a demo with 150 people earlier today which was more on Suno and everything else, so I'm gonna go and tell it to differently.

Currently, actually look at the log from 6 .30 to 7 .40 p .m. today. I think you got a demo from earlier in the day.

Threshold concept he just demonstrated. It'll just take a minute. Sure.

You can pose to the audience now. Based on what you've just seen, from AI creating music, to agents autonomously executing complex tasks, to the utility threshold concept.

Which task in your work do you think is currently sitting just below the utility threshold? What would need to be?

So you can see that it, what it did there is, as I said, it confused my demo firm earlier today. It's asking about the utility threshold and everything else.

So I stop that now. I can actually go and say it again, go here, back to voice.

Retrying with the Right Transcript

Let's place another call, but use the right transcript this time around so that the question actually is relevant to the audience I'm with right now.

We have a little bug and I knew this, that somewhere in the conversation, when you're halfway through and you're using a voice message, it doesn't seem to populate it, but it is actually working in the background, so if I go, I'll have to fix this somewhere tonight, but it is executing on this, so let's see.

Okay, it's not properly doing this. The

The meetup started at 6 p .m. today. I think the transcript will start at 6 .30 p .m.

Somehow there's music coming through there. So in parallel, all of this is still going. So this is now seven minutes in. Let's see where it's at.

Real-World Hiccups: Outages and Lag

Okay, so one of the things that might happen, maybe because my, yeah, okay, because of the, I don't know if anyone had seen earlier today,

there was a big Cloudflare outage, and limitless had a bit of a backlog so what looks to happen is that right now it's not actually uploading in real time normally it's like a five minute lag there seems to be a much bigger lag

which is why it is not able to get the thing but i think you got the general idea the idea is that

Closing Thoughts

you end up with a much easier way of executing on these agents and basically every workflow that you have is just explained in text files that then connects with a whole bunch of different tools and is able to start executing on what you have.

From Chat to Execution

So rather than kind of everything staying in ChatGPT, it just is able to go and execute and do real work.

So on that note, I think I'll leave it there. Thank you.