Introduction

Hi everyone, great to be here.

My name is Johnny Ball.

I am the founder of a consultancy called Axiom

and we help B2B scale -ups with AI transformation.

And tonight I am going to be showing you an agentic system for automatic knowledge and task management.

Okay, so I know we've already done a show of hands, but another quick one just so that I can read the room.

How many people here have worked in a terminal before?

Okay, great.

This is a good start.

And how many people have worked with Claude Code, Cursor, similar?

Great, okay.

So most people in the room.

Good stuff.

That sets the scene.

Tools Overview

Claude Code in the Terminal

So today we are going to be using Claude Code on the right -hand side of this screen,

and we're going to be using Obsidian, which is essentially a notes -taking app, and that's on the left -hand side of the screen.

screen.

So, a quick introduction to Claude Code for those of you who haven't used it.

It is a CLI, so it is an agentic coding tool that lives within the terminal.

So, essentially

used by software engineers to support them with writing code with AI.

However, it is

also extremely useful for other things.

So, both of the tools I'm using tonight, Obsidian

and Claude code were not designed for knowledge management or for managing tasks but they are

extremely effective at it and that's what I'm going to be demonstrating so I'm gonna have to

dive straight in because we haven't got much time and I'm actually going to be running some some

agents live and I want to make sure I don't preach Adrian's time rules and we miss the pizza so first

of all we're just going to launch Claude code straight away to get started which we can do in

the terminal by just doing this and this really for those of you who haven't used

a tool like this before although it might look intimidating for the non

software engineers in the room you can just think about it as chat GPT and with

a different UI so a different interface to interact but this here when we

interact with it it's all natural language it's not quite as user friendly

in terms of the skin as using the chat GPT interface but it's essentially the

same thing there's a lot of extra things claude code can do that other sort of assistants cannot

but for this demo just think of it as chat gpt that can interact with files on your machine

Obsidian and Vault Structure

and here on the left hand side we've got obsidian which is simply a note taking app and under the

hood of obsidian it's just directories with markdown files which are a type of text file

so all of these are just directories with files in them as you can see so nothing complex

A very quick overview of the file structure I've chosen here.

Below System Demo, this here is the guts of the system.

So in here lives our actual content, so the notes that we are creating and the knowledge that we're working with.

I've chosen a PARA structure, which is, I believe, Diego Forte's sort of note organization structure or knowledge management structure, but you could use anything you want.

Inbox-First Capture (PARA and Scratchpad)

the important bit is the inbox which is in this system a capturing place it is a

dumping ground think of it as the backlog this is where you put all your

random notes so here in the scratch pad we have just some some random notes it's

all this is all made up for the demo by the way I'm not real notes that you can

scribble throughout the day and you can also dump files here so this is some

discovery call notes following a trans that could come from a transcript from a

client meeting so current challenges what they're looking for specific use

cases budgets etc but the idea is you dump all of your thoughts and notes in

that the inbox and so we're going to dive straight into using Claude code so

Agentic System Architecture

what we've built or I've built here is an agent operating system and if I run

that code so all I did there was run a shortcut that reads one of the files

that lives in this system demo area so the second part of our system on the

left is the brain this is what runs the agent system but don't be intimidated by

the fact that it's it's doing all sorts of clever things it's actually also just

a load of text files so there isn't code here this is all just natural language

prompts that are written for this agent system and the way that Claude code or

this structure works is that when I type in this here agents demo orchestrator

Orchestrator, Personas, and Governance

All it does is call this file, which is called orchestrator agent, and it's just a markdown file, a text file, written in natural language.

It's written in a very specific way, and I've borrowed here heavily from Lawrence Liu's system, which is an open source system for agents called BMAD.

So I've taken his sort of thought process and borrowed from that heavily.

But all this does is define what the agent is.

We give it a persona, we give it capabilities, and we give it a sort of objective.

and then it can call all sorts of other things from these other files here but

it's simple stuff like templates so how would you lay out a client report it's

things that data which tell you how you should work the constitutional

governance of this system and then the important one tasks which is essentially

Tasks, Workflows, and Activation

workflows so these can be executed by the agent so they get called when we

activate this agent so what we've done here is activate the orchestrator agent

and you can see it's now talking to me it says here we go hello I'm Terry or

demo orchestrator and Terry's job is to help me choose which agent I need to

The Three Agents: Architect, Vault Processor, Task Executor

work with because this is a demo we only have three agents we have the architect

who actually helped me build this system so this system is all built by the

architect of the system is bootstrapped up very meta and then we have the volt

processor and that person's job is to essentially arrange all of the the

knowledge from the inbox where we're dumping everything so put tasks in the

right place to organize files etc and then we have the task task executor who

can actually do things so because we're rushed for time I'm gonna go straight in

and I'm gonna ask the agent vault processor to jump in I could have said

to Terry there hi Terry I want to process and process all my documents and

Terry would have said you should talk to Maggie who's your vault processor and so

Live Demo: Processing the Inbox

So here we go, hello, I'm Maggie.

So what Maggie's purpose in life is, as I've suggested,

is to go into the inbox and to have a look at everything.

Now this is a demo, so there's only one file,

which is this mock transcript with Magic Robots AI

and this COO here, Sarah.

And then there's a scratch pad

with a couple of bits and pieces on it.

Draft up recommendations from Sarah,

pull some quick research,

use the discovery call transcript, et cetera.

Demo Inputs and Context

Then as you all know, Gemini 3 dropped two days ago,

and so I've said you need to pull this YouTube transcript

and write up what's there and schedule a follow -up.

There's one other thing that sits outside of the system.

I like to use Todoist.

This is purely a UI thing.

I like the Todoist interface on my phone.

I like to be able to go like that and pop up a Todoist thing on my Mac,

but I often keep notes in here.

So I'm just showing you that so that you're aware of what's in there.

This is going to be part of what's coming up next.

Vault Processor (Maggie): Capabilities and Commands

So here's Maggie, and Maggie also tells you what commands she has.

So she can process if I give her a process command,

and she can do a quick process which is only high confidence items and she tell

introduces herself I process diverse inbox content etc etc so I've just

explained and she tells you exactly what she does and what she doesn't do how can

I help you today and say hi Maggie I need to process my whole inbox how do I

do this actually that'll just take more time so could you execute this please

Running Maggie: Routing, Extraction, and Syncs

okay so Maggie's now firing up and what all that's happening under the hood here

as I say nothing complicated this here vault executor is just a text a markdown file it's

being called and Maggie has the ability by calling these other scripts in here to perform certain

tasks of which one of the tasks is most likely extract tasks process inbox and root content so

Maggie will be calling these files and just running them and you could literally copy and

season to chat to EBT it's just text so off she goes doing this shouldn't take

long so I've included a little bit of code in this which is not essential at

Todoist Sync: Example of Simple Automation

all for the system but might be interesting I have built an API sync

with to do is so that it extracts my notes from to do is automatically and

marksman is done and you will notice hopefully that the sync has happened so

So Maggie has gone to my Todoist, pulled all of the items from Todoist, put them in my scratch pad, which is my notebook, and has pulled them out of Todoist and marked them as done.

So that's just a little experiment to show that the agent is smart enough to take a script that lives within the system and execute it automatically.

There's no MCP server, there's no fancy stuff, it's literally just a Python script.

An MCP server is a model context protocol, which is a method for large language models

to interact with a code base or a server.

We don't use them in this presentation.

I shouldn't have used the acronym.

Sorry, Adrian.

Local vs Cloud: What Runs Where

Is it running on your laptop?

Everything is entirely local.

Yeah.

Well, no, sorry.

That's a complete lie.

Sorry.

Clawed code runs on the anthropic servers.

So this is not run from my machine.

This is hitting their APIs, and they're going off and using their servers to run all of the smart parts on the right.

But everything else is saved on my machine, so all the notes are on my machine.

Okay, so we've told Maggie to get going.

She's pulled everything into the scratch pad, and now she's just giving us an overview of what's there.

Great, here's what I found in your inbox, and she's going to give me a readout.

And hopefully, so she's also said, before I start routing everything, any additional context.

So, at this point, if we were doing this for real, and it wasn't a demo, I would start

dictating like, well, this file should probably go there, here's a bit of extra context.

And it tells Maggie all the things she needs to know that wouldn't be in the files.

But I'm just going to say go, which is a lot quicker.

And hopefully Maggie's going to be able to work it all out from what we have here.

Automated Filing and Task Triage

But essentially what we're hoping for is that Maggie will understand that this here, draft

up recommendations for Sarah which is related to this the discovery transcript

she's gonna hopefully understand sorry question yes five minute warning shit

already okay she's hopefully gonna understand what she needs to do with

these and I will process them straight away as we're running out of time so

she's basically given us recommendations here of where things should go so she

She said the transcript should probably go in this project's client file, which it should.

She said that we should make a note of this task in the agent tasks, which is the file where we keep tasks that are going to be executed specifically by the system itself.

And then she said these other tasks, we probably just want to put them in the backlog or in this week.

So they're tasks that would be completed by me as the human.

And hopefully she's going away and doing this work.

And what we're expecting to see, you can see the files missing from here.

here.

That's because Maggie has now filed that into a new directory called client -magic -robots

and she's moved this here.

So obviously in the demo this is just one file moving but

you can imagine you could have lots and lots of files that you've accumulated through the

week and you want the agent to just file all of your meeting notes and transcripts in the

right place.

Building Context-Rich Knowledge Bases

The point being you're building a rich context here.

This is the whole knowledge

base of everything you're doing in your work and then the agents have that context in what

what they're doing so she's beavering away and what we're expecting again is

for her to empty the scratch pad which she's done great she's now populated my

tasks these are human tasks she's prioritized because she's noticed the

Human Tasks vs Agent Tasks

dates and she's put today as the mind stone talk I need to prep for this week

do these other things and then in the backlog because it's q1 she's understood

she needs to put the land graph research there and then in agent tasks she's

taken the agent tasks she's enriched them she's tagged the context she's put

put the actions, and we're going to execute these straight away because otherwise we'll

run out of time.

She's put the context and she's essentially told the system what it

needs to do.

On the second task she said, you need to go and pull this YouTube transcript

and analyze it and write a summary of what is going on.

Live Demo: Executing Agent Tasks

Hopefully, as Charlie, our task

executor loads up, we're going to skip straight through and it's the same process.

We're just

execute all and he's hopefully going to just go through and look at these agent tasks and

actually start doing them.

Let's see what he says.

Okay, excellent.

He's found the two

agent tasks.

He's going to create an execution plan for them.

Planning and Parallel Execution

He's pulling the templates.

So you can see here, it takes a little bit of time for the system to do the thinking

but you can imagine actually doing this manually if you wanted to go away and research, look

look through the transcript notes that you'd written,

go and collect relevant context.

I'm just gonna say yes to keep it going.

So he's just asking us here.

Here's the execution plan.

I'm gonna fetch this YouTube transcript.

It's actually gonna use another little bash tool I made,

but that's not relevant.

And then it says he's created success criteria

and he's going away.

And you can see that he's doing web research for the task,

conducting all of this.

He's running them all in parallel to be efficient.

Blending External Research with Internal Context

And hopefully, what we're going to get out of the end of it is this fully written up research that is not only using external data, but should be pulling as well from the specific context that is unique internal data from this client discovery call.

Now, whilst that's running, any questions?

Q&A and Best Practices

Do you have a GitHub or where we can find examples of something like this?

Resources and Inspirations

i do i do have a github but this is not on my github but i would i would recommend to take a

look at lawrence liu's stuff which um as i said it's called be mad and i've heavily borrowed from

that in this demo my actual system deviates a fair bit um you can look also on tiego forte who does

second brain he's done a lot of this knowledge based work and then the subreddit on uh knowledge

based systems or similar if you find the subreddits on it there's a lot of chatter there that you can

you can borrow from that's exactly it that's it a couple of days i mean to be honest most of the

thinking so in my real system i don't follow the para framework anymore so it doesn't scale as much

as i would like most of the thinking was the architecture and the whole principle behind it

like working out how i can take the bmad framework and apply some of those lessons of modular agent

work and like lazy loading for the context so working it all out took the time actually

Bootstrapping Agents and Constitutions

executing is all bootstrapped the first thing I did was build an agent to build

agents and then I bootstrapped it all up so now once you've you've built the

builder and then you put in place critically which I've not covered here

you put in place a constitution these are the rules of the vault it's like

like a major linting guide if you're a software developer that anything the

agents build have to follow the Constitution and then that keeps the

whole thing in order and that's that makes it easier to bootstrap I'm just

going to load up the fight the one file as we've done it and then I'll jump on

Results: Fetched Transcripts and Summaries

other questions so they fetched the transcript successfully I'm hoping

they've put it somewhere sensible maybe resources what does it tell me here no

oh yeah resources research notes okay so it's put it in research notes so you can

see it has just pulled the entire transcript from YouTube that I

recommended that was actually a script but it could have equally done it itself

it's just faster to do it with the script and it's going to use that to now

Generated Outputs: Gemini 3 Summary

write another file there we go so we now have from that transcript a Gemini 3

summary which is giving us an overview of the Gemini 3 launch the key features

and performance metrics coding capabilities etc etc etc so a fully

detailed very long review of Gemini 3 based on a model that was only released

two days ago and a very massively written note in my to -do list we end up

with this report in, well, I'm over time, in 15 minutes.

So whilst this loads, let's

hit that.

Let's do some more questions.

Is the agent that's supposed to do this, is an agent?

Yeah.

So that's all codified in the agent files themselves.

So that's the task, essentially.

When you're building these agent systems, it's all about making sure that the agent

does what you want it to do and trying to capture that in the prompt.

So that is, yeah,

it's a big part of the architecture.

got it so the question was around how you keep the the system honest essentially and you can

how do you use the constitution to make sure that the architecture you're building is appropriate

and the agents know what to do um within the agent files there's very strict guidance about

using the Constitution and other parts of this such as the the workflows

themselves will be really explicit like very clear about what is required so the

input tasks required what this task does the execution steps and these are all

essentially very strict guard railed processes it needs to go through and

it's lazily loading the context so when it needs more information within the

agent prompt it knows to go and get that specific information and only load that

So you're trying to keep the context sufficiently compact that it doesn't drift off and get confused.

So it's only lazily loading what it needs.

It's a modular agentic system, which is different to the other way to do that.

Another way to do this is have AI playbooks, which is what I also do, where everything is just in one file.

But what I found when I scaled that system is it becomes really difficult to stay on top of all of your different playbooks and know which one's for what.

And they're all quite rigid.

whereas this you can chat with the agent and say I've got this this idea I want

to do this and talk to it and it's still constrained in what it knows its job is

and you can work out work out whatever you want really I'm just going to

Generated Outputs: Client Discovery to Report

quickly show this so task to quickly we have loaded here so we had the client

discovery notes that we started with about the call we had and now it has gone

and written this client report for us which is a fully detailed report and it

it is bespoke for magic robots based on this discovery file here but it's also

gone out and done a huge amount of landscaping industry land escaping key

trends competitive intelligence and it will have made this all fit perfectly to

what is included in the original discovery school so you've then got a

report for a client that's bespoke and it's done deep research combined it all

together all again just from a quick to -do item within your scratchpad that's

Conclusion

the demo over I'm not sure how I'm doing for time but happy to take questions if there are any