Have you ever wanted to be the director in your own The Truman Show? Introducing scenarios to see how your subjects would react? Well, if you have I may be able to help you there.

Earlier this month, an article caught my attention, regarding an experiment that was run asking 4 agents to create an online radio station for 6 months. The results were fascinating and scary at the same time. I wondered though, could I build a similar simulation that would mimic the C-level of a private equity owned company?

What sort of mayhem could come from that?

While showering one morning (tis where I do all my best renumerating), I thought through how I would do this. I wanted to assign each C-level their own model, their own backstory (system context), and their behavioral traits. Given that I wanted them to interact with one another, the best environment to set this would be a Slack channel, where we could easily observe the output. That solved the UX issue.

I then wanted them to be able to interact with each other, in a random, but yet predictable manner. In that, we know for any given channel of people, not everyone responds to every message, or they will lag behind – yes there is always that one person that answers that message from last week, completely out of context from the current one, and confuses the shit out of everyone else. You know who you are.

But if I was truly to take on Ed Harris in The Truman Show, I needed a way to inject chaos. This chaos could be a simple event, such as the weekend, site going down, or some awkward HR incident. That would be fun. Though I could do better. What if they could respond to real live news events – dropping in a URL to a news story and let each of them read it, respond accordingly.

So I gave my “C” level agents, the ability to do web searches, as well reading the text from a URL. That added a whole dynamic – but again, I wanted more. cue evil laugh What if my little virtual C level executives had a sense of humor (well someone of them at least) and they could express their feelings through the medium of memes?

So I gave them that ability too.

It was all starting to come together very nicely (though the water was starting to run cold), and I thought memory needed to be handled properly. We need them to “learn” about the people they work with, remember particular interactions. Carefully managing their context, and summarizing it accordingly, I built-in the ability to inject into the system prompt the characteristics of each of the people in the channel from their point of view of them talking to them.

After drying off, grabbing a green tea (naturally), I got to work on coding up this little experiment. Through the power of OpenRouter.ai I had access to pretty much every model through a single API call. This allowed me to let the personalities of each C level be powered by completely different beasts.

Working through some of the timing issues, creating the necessary tools for them to consume, I got this up and running, and initial results were fun. I sat back, and just watched the drama play out. I tossed in the odd Reddit story to see what each of them would make of it.

I got bored of looking at just their names, so I toodled off to Gemini and asked it to create me random profile photos of each of our people and tossed them in. Ooh what a difference a little 48×48 image can make. As I was reading their output, I thought it needed one more feature – the ability to have side 1on1 chats with each other, so they can talk about others behind their back. Ooh that was the missing piece.

I have worked with some wonderful (and not so wonderful) characters in my time, so it was intriguing setting up those personality types, and then letting it run. What surprised me, was the AI was nailing the responses that I would have expected. It was spooky – or was it just confirmation that cliches do have origins.

If you want to have a go at playing Ed Harris in your own Truman Show, then head over and clone my GitHub project to start having some fun with Collin Canning (bonus points for who knows that movie reference)

https://github.com/a1anw2/collincanning

You create the context of the company they work for, add in as many people as you want into the chatroom, give all completely individual personalities and assign them their own AI model. Once it is running, you can then through the admin UX, poke in events or give them a URL to go and look up.

Be warned – it can be addictive and highly amusing.

Be sure to check out my piece on how prompts actually work under the covers, this will give you greater insight on how this was achieved.

AI Disclaimer: Gemini Nano Banana Pro was used to generate the photo – The Truman Show