Introduction

One thing I wanted to touch on, to follow up on, is the use of AI agents and where we're moving to in terms of using AI in our workflows.

AI Agents vs. Standard Chatbots

So everyone's pretty comfortable with using LLMs and chatbots by now but what we're seeing is people that have the ability to customize their lifestyle by AI agents are just a bit more productive than using just usual chatbots. So

essentially the main difference for those who aren't familiar is an AI agent is basically just building on top of an LLM. So let's say you have a language model like Claude or ChatGPT,

instead of just using it as a chatbot where you wait for it to respond and it gives you answers, you basically give it tooling and it basically does things for you.

What Agents Can Do With Tools (Email, Web, Calendar)

So for example, if you give it the ability to access your emails, then it can send emails and respond to emails for you.

If you give access to the web, it can search for you.

If you give access to your calendar, it can book meetings for you and stuff like that. that.

Why the Chatbot–Tool–Agent Line Is Blurring

One thing I will say is the more advanced the LLMs get, it kind of blurs the line between what is a chatbot and what is a tool and what is an agent because a lot of these LLMs now have tools like they can access the web, they can access your calendar if you give them permission and stuff like that.

Using Multi‑Agent Systems to Simulate Real‑World Scenarios

So I basically had an idea a few weeks ago about using a a bunch of um agents to basically simulate real world scenarios and see if we can use that to extrapolate predictions so uh this it does look a bit confusing but when i explain it'll be fine so

think of one um ai agent as like a specific so let's say you wanted to simulate um a population of i don't know uh a hundred thousand people in a certain environment and you wanted to kind of figure out depending on the action these people take who would they i don't know vote for in the

next general election essentially and what this tool does it simulates each individual as like a an ai agent and then it gives them access to certain information and context and tools and it allows them to make decisions based on certain stimuli and what i mean by stimuli is

for example if the government announced a new i don't know law you can kind of see how the people would react with that so it's basically just the tool to kind of predict real world scenario so

Event Streams, Seeds, and Ontologies

under the hood this is basically the event stream so you can see that the system starts in basically works in a series of states 1so you essentially start at a null state where no AI agent has context of anything and remembers anything and they just start up you then then basically get given a series of context.

So for example, agent A could be a policy maker, agent B could be someone in the media, and agent C could be just someone in the public,

and they'll be given information based on how you've pre -configured them. them.

So essentially this tool, it allows you to basically enter in some sort of information. Once you've built your foundation, it basically gives you enter in a certain stimulus to kind of see how your ontology reacts to that.

1So you start off at seed and ontology. Basically, this is just documents or information that you give it, and then it builds like a graph I'll show you the graph in a little bit.

It basically builds a graph, which basically maps how all the different agents kind of work together. And this is basically the language that agents use to understand what's happening in the real world.

So it creates all the agents and then it connects them together in some sort of graph. And then every time something happens, so for example, if you upload a document of a new policy,

the agents all have a collective understanding of their role in the graph which simulates a specific scenario and then they can act according to how they should be acting in accordance to the wider graph.

Why Local Agent Context Matters

One positive about using this specific technique is that each agent is local, so essentially So actually, each agent has context about itself and not, not, so for example, it doesn't have like access to like,

so for example, if I asked Claude to predict something for a certain outcome, it has access to like the internet, it has access to like all this pre -trained data. So you're not really sure if the prediction is giving you is something that it's accurate.

Well, not accurate, but you don't know for sure if it's something that it's predetermined from what you've created or if it's something that understands without you knowing essentially so giving making sure that all the agents are local makes it so that they can't basically take a take a peek at what is happening outside of their specific context which is kind of like spoil the the entire experiment so moving on so the graph

would basically look something like this so essentially it's one large one large graph each node acts as a specific category of an individual so in that

scenario where there's a population of people and you want for example like let's say I pick this room out for example some people here are I don't know bankers some people are construction workers some people are

policymakers everyone's different but once they've been given all the information they've needed, you then run a series of rounds to basically see how they react to certain things.

So in the

Example Walkthrough: Simulating a Middle East Conflict

example that I'm going to be showing you, I decided to kind of predict the outcome of the conflict in the Middle East.

So a few months, like a month ago, essentially, there was like a conflict in the Middle East with like Iran and US. Yeah, yeah.

I kind of had a feeling it would be difficult to predict because I'm predicting something in the future. I could have gone a bit easier and predicted something that's already happened and then kind of backdated it, but I was like, you know what, who cares.

You know what, I want to show you guys how the graph looks. I think I have it here. here. Yeah.

The Ontology Graph (Nodes and Relationships)

So when you once you run the first round, it basically creates this is this is something called ontology, which it basically a series of nodes that will understand how they connect together together.

If you've, if you've done, this is kind of niche, but if you've done SQL, and you understand relational databases, it's kind of like that.

So yeah, so for example, each, each node here is given a certain specific information. So if I don't know if I select one, like, what is this. So for example, this is a specific tanker in the Middle East carrying oil.

If you go to some of the larger ones, you can see why they're there. So for example, here, Israel, I mean, they're quite important. You can basically see how everything is interconnected together. together.

It does also depend on the model you run. If you have access to more compute, you can create larger graphs.

Some guy in China did something similar to this, and he had nearly a million nodes. That level of compute is insane, which I didn't have access to, but this is still pretty decent.

I was quite pleased with how this came out.

Now it's time to run the tool. The problem is the tool takes quite a while. It takes like 10 minutes to load up, which I couldn't do. So I basically ran the tool already, which is cheating, but I'm sorry.

Running the Simulation in Rounds (Day-by-Day Updates)

This is basically how the tool looks. So I basically configured this, I gave it 10 days, 15 days, 15 days of basically 15 rounds which simulate a day.

So on on day one, on day one, all the main actors are basically turned on. So they're at like state zero, T zero, basically. And then something happens. And then the entire graph updates in terms of the information everyone has, and they

act upon that information, and then something else happens. And then it kind of like keeps going, basically.

so you can see so something you can kind of see so at the beginning you can see

a de -escalation so Iran tried to de -escalate the conflict early on

I probably should have told you this but I started this from February 26th to 15 days which is I think March 13th

mid to late February is basically kind of when we started hearing murmurs about maybe there might be a conflict and stuff like that which is when I started it so

so the early seed rounds you can see Iran trying to de -escalate and then you can kind of see sanctions so before the conflict started the US was going heavy on sanctions and limited military actions to basically get involved but not directly get involved and then obviously

we can see limited military action from Israel obviously we know it was Israel that like launched the first you know kind of like started the war off i'm i'm not hey i'm just i'm just reading the facts here uh so you can see here uh the zion entity continues his aggression yeah yeah

so iran basically threatened to close the strait which which we know later on they did close the strait but that was after the 15 -day period so you know i would have been cool if i could actually put it that way uh de -escalation from from saudi arabia obviously they're oil so they're like like really implicating this.

So as this has happened, oh, what has happened today? So as this is happening, every, so at the top you can see like day 15 out of 15. If I, I think I have a little recording.

Okay, something similar to this, right? So what you could see is like every time a new day would start and there'd be an update, the graph would kind of reflect depending on the predictions that it's made.

So across the 15 day period, the price of oil started at like, a well the model guess that it started around 55 which is close started like 60 and then it

presumed to ended that right i don't even know what that is maybe like 94 or something like that so let me go back to this guy here so uh okay um i know it's fine

Automating the seeds, or is it a manual mod? Automating what? The seeds.

Are you getting the updates going? No, I'm building one large graph, and each node is an agent, so underneath the agent is a language model. So it has its own reasoning, essentially.

So this is kind of small, but let's get to the ...

Results and What the Model Predicted Correctly

Okay, so what the model predicted accurately. The model accurately predicted a split between the rounds.

so the pragmatists and the hardliners so there are people on both sides in around both sides that wanted de -escalation and then there were also people who uh kind of wanted to like ramp it ramp it up um us relying mainly on sanctions and military presence rather than jumping straight into the war uh but then then they did jump straight into the water that um israel using targeted limited strikes to manage threats without full escalation uh and then saudi staying neutral and focused on keeping yeah yeah so these are things that the model predicted accurately um from

the 15 days that I gave it so something more can you guys see that I hope you can see it so on the far right you can see the the prediction I showed you and then in the in the the two graphs are the prices of oil so as you can see

yeah so you can see when we started we started February 12th and the price of oil was I think that says 62 or 67 our model predicted that will start around uh started around 55 euros and then the final price after the 15 days was 92 dollars and our model predicted 94 dollars so it was actually quite close

Practical Uses, Limitations, and Warnings

so yeah so basically what what you should be using this tool for so basically you should be using this tool to like rehearse decisions of uncertainty it's not something where you shouldn't be using

this to bet on things basically so some of the some of the things that you could use it for like are planning like scenario planning for markets and geopolitics that's what we did modeling

customer reactions for new products so if you're a business and you want to release something and and you're not sure how the let's say you're like jaguar and like you had that whole thing about releasing your new cars that look really weird you could kind of do you can kind of use something like this to basically feel and figure out like oh people hate our hair products um and then yeah

just exploring exploring what ifs it shouldn't be used to determine um things so for example like authoritative predictions um or compliance you should you should not be using it in scenarios in which if you get it wrong you'll be in big trouble and it shouldn't be used as a trading tool as well.

You shouldn't put money on AI agents.

If you need something that's reproducible, you also shouldn't use this.

Solid, man.

Conclusion and Next Steps

Yeah, that's about it, man.

If anyone wants access to the tool, they can use it.

The code is kind of messy.

If anyone has any questions, you can ask me.

Do I still have time?