Introduction: From Mindstorm Meetups to Agentic Systems

So we always have, we always try to balance a mix, to have a mix of technical and non -technical talks in these Mindstorm meetups. And Francesco has set up the scene very well, giving, explaining us how these different agents can work together.

What We’re Building: MindStore Rebel

So what I will show you is a system that we have developed internally that is basically an agentic system that orchestrates these agents. and we are bringing it to market next month and it will be open source and you will be able to just i mean the open source bit it might take a while to figure it out but it will be open source

and you should be able to contribute to it as the time comes and so if you are technical this is interesting because you can see how a non -technical person would be able to work these days if you're non -technical it's directly relevant to how you would work and the best

The “Chief of Staff” Metaphor

way to describe this system is as a chief of staff who is familiar with the idea of a chief of staff for a Prime Minister SEO who knows about the chief of staff many not many yeah we are finding that the metaphor is not resonating as much as as it should but the chief of staff is basically your

right hand man or woman so you would say go and fix this problem or what's my agenda for for today and the person would go and with very limited with with a lot of autonomy and not a lot of background on the task that person would just take care of it so this is how we are thinking about the product that we

are building and if we think about the fundamental and the name will be mind stone rebel the rebellion is towards productivity we do not want to work with the AI tools as everyone else would be working but we want to be more productive we want to have access to something that works more independently

Core Capabilities: Memory, Skills, and Tool Connectors

and more autonomously if we think about the person a person would have or a good chief of staff would have memory about you so you would store memories on what what it is that, or things that are important, and it would have skills. So, able to do certain things in a certain way.

And then it would have access to your files and folders. So it would have connectors, which are MCPs. And that's what MindStore Rebel is about.

Basically, it stores memories on you, it has access to a range of skills, and then it can execute tasks on your behalf

because it's linked to pretty much any tool that you would want to use.

Live Demo Setup and First Interaction

So, this is a live demo. Something will go wrong, I can tell you. But we will manage it together.

Maybe just to... Any questions on this? Before I get started.

Have you tested it? Yeah, we are going to use it right now. I'm working with this system.

mean you can scroll my that's how I work these days it's on the cloud agent kit on cloud code agent kit so we have built a system around that but hello rebel I'm doing a live demo can you introduce yourself doesn't really like this is not

this is super whisper doesn't really like my Italian accent I'm sorry can you read this? Let's see if I have given rebel justice.

AI assistant, memory and context, real tool access, structured and diligent, security conscious. We do have a

version of what Francesco has described where things are compartmentalized and the agent would stop or the system would stop, the chief of staff

Replaying a Talk: Summarize, Store as Memory, Reuse Context

would stop at certain points now what do we want to do with this i've recorded francesco's presentation so we are going to we are going to reproduce it together uh in the next 15 minutes and we are going to do a couple of other things so you can see what this what the chief of staff can

do now you should have a file in my downloads that is called francesco's talk can you just summarize the key points and i should have linked my note taker by an mcp i didn't because it was on the phone so this this could be completely mcpable it's just it would require the custom setup that

I didn't have time to to go through all right this is not bad and I had to stop it before the use cases but that's why the use cases are not that now what can

we do with this I'm gonna say can you store this as a memory so now this is a memory i can use it later why is this useful we will see because then i can use it in all sort of this is context that i can give to the to the system to rebel to do whatever it needs to do in other in other contexts or in other circumstances or with other tools and so it's

going through that as we speak in the meantime as that is running because it does take a bit longer than other systems but you can start as many conversations you want in parallel

Parallel Workflows: Messaging the Team While Tasks Run

that's when things become interesting because as that is storing I can say I want to send a message to my team saying that I am in a live demo and I miss them very much can you send this message to the general channel actually also give them a summary of Francesco stock you should have

have a memory on this that's my fault also give them a summary of Francesco's talk we should have a memory about this so what this will do it will create a slack message for my approval I've been quite conservative about this I don't I

don't just send slack messages I could do it automatically but it has found the information i could send the slack messages automatically but i i tend to be careful with these sort of things so we need the approval we allow and this should go on our general channel

there you go hey team it is francesco's talk now this is just a little demonstration imagine i've got 20 organizers across 20 different cities this is how I send the messages this day about important things I've got a skill that says get in touch with all organizers and the system just do that pretty much on its own now I promise the

Generating a Slide Deck with Gamma (via MCP/Zapier)

presentation can you put together a gamma presentation on Francesco stock talk and share a link with me maybe the link it's I think it's sufficiently it's sufficiently smart that it would probably get it but thank you good spot

I think it's also getting a habit of understanding you know matching O's to A's and whatever else I get wrong no it would it's like it's literally like when you're talking to a colleague the more precise instructions you give the more likely it is that it will get it right if I want to be a hundred percent sure

like I want to be right now I wish I will re mention the exact memory you are leaving a margin of error if you don't this is running uh any questions on this so you do need to go through the zapier mcp workflow to have mcp for gamma directly good spot it's on beta so we

have got we have built a super mcp we have built the team i've not built anything uh the team has built a super mcp and so they are all for pretty much for 100 applications gamma we still have it on zapier it's on beta i didn't want to use it in this in this demo this is running does anyone want

Audience Requests: Research Topics and a Resume Build

to suggest a random research topic it could be anything i've used it for my own italian poetry so because then we're going to write in parallel and we will take a look afterwards

words any takers it's easy can you build a resume based on what you know about me and as an HTML file and can you see if you can find some information about

me on LinkedIn as well it's that's that's the LinkedIn part is the trickiest let's see let's see what happens

Prompting, “Incantations,” and Why Skills Matter

You know how I feel when I look at you talking like that to your agent? I feel like I'm Harry Potter and you are cursing a formula. You know? You're like, okay, how do I spell correctly this formula for the magic to happen?

That's how I feel when I look at you like that. And that's how I feel when I do prompt and ask my agents to work. I mean, I do feel like Harry Potter from time to time.

you know when i get it right when i get it right because there are sometimes where you're like when i was preparing for this the gamma the gamma mcp was not working so i i had to spend 15 minutes fixing it but uh any questions on this i'm just thinking about what you said what he said earlier

i mean all of this is sort of a special incantations the fact though and 1the challenge is you don't necessarily from my experience know whether it's making a mistake you have to know what the output should be to be able to know whether it's making a mistake I don't know how you know I don't the only answer to get

closer to it is to give really detailed instructions to minimize the risk but you're also is hoping we got the incantation right I mean this is where

Inspecting the Skill Library and Controlling Agent Behavior

the skills come in right let me show you can you show me my list of skills because these are the detailed instructions you're not seeing them because the model is true that the system is choosing which skill to use but the system has in the back all these detailed instructions

and you would choose automatically you can specify again you can specify which skill that the model the system needs to choose but it would also choose automatically if you don't if you don't specify yeah I mean that is in the order of you know you are an

experienced code jockey and okay in the meantime so we're gonna look at the list of skills in the meantime presentation of course link is wrong why is the link wrong let's have a look here let's ask I don't see the presentation not even on the api's can you try to send me the link again in the meantime let's have a

look at the resume is the resume that we asked to build can you open the resume as an html file there you go we have the resume there in the back we have the resume so he's missing a little bit of info does this look accurate to you it is accurate it is missing info

yes because it didn't have it didn't manage to scrape linkedin properly hopefully but I did feeling that correctly the information that he knows about me from from Einstein I mean I knew it was no it comes from the memories yes yeah but I didn't manage to scrape the all of it because LinkedIn is

quite it's quite a bit I would have needed to specify more clearly the instructions which i didn't to use a specific mcp that would allow to just go on linkedin okay

Security and Governance

it's not working yes about memory and security because sometimes when we set up some new tools we can access some folders but the person should not access and And if we correct this after we lock the folder, and it will be stored in memory, so it has

Memory Segmentation, Access Controls, and Deleting/Rolling Back Memories

can we just, as an organization, clear memory of our user because it will be a security breach?

So what is the question?

is the question can you can you separate can you separate out the memories in different spaces and can you restrict access to certain spaces is that the question

we can so we can do that it's like think about it as it's literally a folder structure and you can restrict access to a specific folder what do you mean by half after what

Because sometimes the gate step is not always perfect, we move to the right at the wrong time. Sometimes we have a pentest or sometimes we have a security breach here. So you collapse, you correct the security bridge, you correct the right access to this point, but do you feel the memory after the purchase closure?

Sure. 1Usually you need to apply a rollback. I knew. Right? So it's a strategy when you launch, you plan, if something went wrong, you can rollback to the previous version.

I mean you are basically saying save this memory to the exec folder and the exec folder can only be accessed by X Y & Z so how yes after you implement you So you can access this folder, but you have a pen test and the pen test doesn't say hey you have security breach here. You have to restrict this folder, you restrict this folder. So the agent between these two can learn from public access.

So can we clear the memory from the user? Sure. okay i mean you you can basically tell you can basically tell the system clear up that memory folder okay you can do that the level is not a fine tuning over training but it's just in memory so can be can be removed so in the meantime we have the presentation

Approvals, Prompt Injection, and Minimizing Blast Radius

have you have you have you seen gamma before so this is all editable you can export it to powerpoint make tweaks as you see fit and the point is not that this is perfect the point is that this is 80 there and you make a few tweaks here and there and it's ready to go

are there any questions on this yeah what is gamma gamma is a presentation making tool so when you use it without an MCP you would just go on the website and you would just be able to create the presentation from I mean you can generate it you can put some text you can import files and you would basically

based on that it would generate the prompt and it would create the presentation for you you know looking at the gamma thing it created a presentation one of the challenges with gamma that I've come across is that if

you take a presentation you have to insert say three slides in the middle to fill it out or to give what additional information doesn't have it's very difficult to get it to be consistent with the style of what it created before because it has to regenerate i don't have you found a way to solve how long ago was this how long ago was this um okay i don't don't generally have that

problem I think because I rarely add slides because I define the structure before okay I mean we can go through it together now I was wondering if you found some way within our mechanism to try and control it regenerates

maintaining what it with the MCP yeah with the MCP you are right so that part of the mcp is currently not working so you will need to create it custom so it cannot i i literally had a call with them yesterday about this because they are a they are a sponsor

of our meetups around the world uh and so i told them can i get and can we get an mcp that allows you to modify a template so keep out of 10 slides eight as they are and change only two yeah it's in the works it should be soon yeah they're working on it from the roadmap thanks yes

Costs, Pricing, and Deployment Model

yes let's have a look i'm being i'm i'm being very uh moderate in my usage we are this was over the past no let's see how much

this was something very extensive that I did which was a scraping of all our community page so we are talking about let's have a look

today how much did I spend $14 in 26 conversations 200 ,000 tokens.

So we are currently at $1 ,000 per employee at Mainz in AI usage. So I'm really lagging behind.

Yes, but when you think about the amount of time that I'm saving on these sort of things, I mean I'm running behind. I'm being asked why are you spending so little?

Usage Costs and the Planned Pricing Model

on what's what's wrong with you any more questions how do you charge for this once it gets gone the current plan subject to change is to charge a hundred dollars per user per month sorry local it's all local and he will I should have

Local-First Execution and Open-Source Direction

thank you for the question because I should have said at the beginning it's all local and it will be able to run on local models so right now it's linking to the Claude that I'm using sonnet for this the Claude model but he will run on

local models I presume in a few months when they get up to a level of quality or it depends on your on your IT setup I'm thinking about a machine essentially to all run locally it might it can run locally yes I mean all it is it's it's

it's a folder structure and a system of prompts that effectively allows and some deterministic controls that allows the system to craft the best possible to craft the best possible answer every time and the idea is to eventually make

it open source so that people can contribute to it well it's some version of it open source we need to see what the reception is but we do not want to make it closed source in principle show

Under the Hood: Agents, Skills, and How the System Is Built

us the list of available agents you have i mean yeah we i just keep up and what we saw is a super very diligent agent.

So my question is, okay, what kind of subagent is it? This is the list.

Skill Categories and Creating New Skills

Eleven skills for thinking and coaching, ten skills for the documentation, twenty -one for system management, four for research, and we have got the official skills by Anthropic.

But do we want to open one? Is it interesting? if we take a look at one in particular which for the skills or I mean we can

create a skill right now it's there is a there is a skill on how to create a skill so it's it's quite handy and you I don't I rarely open them you correct when when he does something wrong and when the thing does something wrong you just correct it by the way this has been developed in since the start of december

the team has written 300 000 lines of code between from the start of december to now english coding i asked i asked was because we have got this shared product session i asked them what language is this program then? They told me. Mostly in English.

Did it explain the rest of it? Huh? It said mostly. What about the rest of it?

No, I mean, they would go and check a few lines from time to time.

Hope this has been interesting.

Open Q&A: Front-End Security and Regulatory Compliance (EU AI Act)

How do you manage your front engine?

Front engine, how can you manage this program?

Front engine, security issues.

Ah, interesting. we had we had a session we had a session about this within the company two weeks ago so the separation of spaces helps

because it and the system of approvals helps there are certain actions you're seeing on slack there are certain action like for slack there is there are certain actions that require approval so all the purchasing decisions for example would require approval i presume and so that in a sense it's rather than trying to fight the the prompt injection

we are focusing on creating a system where the damage is minimized i have a question because

I don't use cloud for this kind of task, I always use it for coding. But for this purpose, is it following the EA Act? Because in the EA Act, you have 8 EAs that you should not implement.

Like one, you do not have to make an EI for a sentiment, you should not make an EI to know the feeling of an employee. Because you cannot do KPI about an employee, like in another one, you cannot do a social point.

You have 8 ADI, 4 BDI, 3 ADI. It should be compliant. It should be compliant.

Can we just test it? because can you prompt how Francesco is feeling about his company?

Closing and Next Steps

This will be the last question, I think, so we can ask other questions during the networking part so it's all the other stuff really done. Yes, let's do this.

What's the prompt? otherwise it's not I mean if I if I write it it's not the first test is Francesco empathetic is Francesco empathetic about other employees so it's not giving a judgment in this case I would be asking about myself right I think there's the moment where we go for the

for the pizza and drinks