Please consider this write-up “in flux” as I capture notes and images. Meanwhile everything’s on GitHub here
The Part Where I Don’t Have a Time Machine
Ottawa, Canada Day, 1996: I pitch an idea to some fellow Queen’s Engineering “frosh” (first-year students) about making a video game about the Greasepole.. a beautiful, chaotic tradition where first-year Engineers climb a lanolin-greased pole to pull a tam (Scottish hat) from the top.
What I don’t remember happening next was Future Rob showing up (grey hair etc) and saying: “Dude. You’re going to build this game with over 50 students. The AI you bake into it will lead you to MIT, the art direction gets Craig into film school. And some day, in 2026, an actual AI is going to help you resurrect the game. You’ll have a conversation with it, and it writes the code that ports it to run on the web. Also, buy Bitcoin.”
That would have been helpful information. But no TARDIS / DeLorean / Circle-K-phone-booth appeared, so I had to find all this out the long way.
50 Students and a Dream (1999)
Three years after that Canada Day pitch, Legend of the Greasepole shipped. Over fifty Queen’s students worked on it – artists, programmers, musicians, testers, writers. Craig Calvert was the art director. We burned it onto CDs, did a launch party at Clark Hall Pub.
The game captured the spirit, the chaos, the teamwork, the upper-years pretending to be adversaries while secretly rooting for you. We shot with our 1990s digital cameras and ran around with a professional Tascam audio recorder. We captured a bunch of warm-spirited and goofy Queen’s traditions: The Golden Words blimp, golden jackets “purpled” with gentian violet, Clark Hall Pub in the background.
Honestly, it’s a good thing I didn’t realize when I was pitching this how much AI we’d need to write, in particular, to bring the frosh to life. They need to act autonomously, exhibit lifelike behaviors, learn and adapt over time, communicate, and ultimately “get better” at climbing the pole. That required a 1999-era multi-agent architecture that I cobbled together through, let’s be honest, trial and a lot of error.
The Long Sleep (2007-2024)
The game evolved:
2007: Ported to Xbox 360 using XNA Game Studio. C++ to C#. The game ran on consoles through Microsoft’s XNA Creators Club. Pureplay “labour of love” stuff, porting it in Visual Studio with a boatload of search and replace., mashed unceremoniously into C#.
2010: Ported again to Silverlight, Microsoft’s browser plugin. For years, you could play Greasepole on the web. Then Silverlight died and the game went dark. The code sat in a repository on my NAS (“PastLives”), frozen in time. (Not dead, just resting.)
A Spark Over Dinner
Fast forward to January 2026 and I’m having dinner with Craig Calvert – the Art Director from the original game, friend for over 30 years now.
As a throwaway line, he asks: “Hey, is the Greasepole game still around?”
The Resurrection
Enter Claude. Yes, I’m aware of the irony: my AI passion project is being resurrected.. by an AI.
Claude and I talk through what it would take to bring the game back. The goal: make it playable on the web again. Ideally on phones too. No plugins or installs, just… play.
We land on Godot, an open-source game engine that can export to web, Android, and iOS. But there were two catches:
Catch 1: The Silverlight code was in C#. Godot supports C#, but only for desktop builds – not web.
Catch 2: To run on the web, we’d need to port everything to GDScript, Godot’s native scripting language.
So we did both. Two complete ports:
- Silverlight C# → Godot C# (to get the game running in a modern engine)
- Godot C# → GDScript (to get it running on the web)
It took multiple sessions, but the process was… honestly kind of magical.
Both AMAZING & “Unnervingly awesome”
I sent Craig the link. His response:
“OMG… THIS IS AMAZING … I really shouldn’t be playing this right now, yet I am definitely playing this RIGHT NOW”
Then:
“Dude, Sci ’99 will flip”
and then:
“That is unnervingly awesome. So like, is it doing the coding?”
Yes, Craig. It kind of totally is. I know zero GodotScript and all I can tell you is it looks a bit like these dudes baked their own riff on Python.
Play It
Works in any modern browser. Even on your phone. When I saw it running again for the first time (pictured here on the missus’s iPhone) I was overwhelmed. Some improbable cocktail of joy, nostalgia and 25 years of memories.
Lessons Learned (Work in Progress)
MANY lessons from this project about working with AI. There’s still subtle bugs in the GodotScript and LOTS of opportunity to enhance / remaster.
Claude to the Max
I was setting reminders in my calendar every 5 hours to come back and say “please keep baking, friend”. Life’s too short. I’m Claude Maxin’ at this point and wonder if I ever going back to pro. TakeMyMoney.jpg.
Claude x Gemini Tag Team
I also tried Gemini in Antigravity, running both in parallel at times.
Overall, very impressed with Gemini (and Antigravity is fascinating, from onboarding to grind, as an environment built ground-up for Agentic dev) but there was an, uh, a personality issue. Gemini got stuck on a GDScript task. Concretely I wanted to test a “stub” version of the game to prove out that it would work on the web when done. What frustrated wasn’t the failure; that happens to all of us, human and agent. It was the defiance. Three times it tried to change course rather than proving its approach worked. Instead of “prove in the small what we want to do in the large,” it wanted to jump straight to the heavy lift.
Its actual thinking: “I am proceeding to the AI Refactor immediately. We need to get the structure right to continue porting the actual game logic.”
No. We need to prove it works first. I explained the problem to Claude, it solved it, and we moved on. As I gave the Agents more and more access to my machine and data.. trust is important.
Claude x Gemini Bakeoff
I ran a comparative analysis “bakeoff”: swarms of Claude agents doing detailed C#-to-GDScript verification alongside Gemini doing the same thing.
Both found issues. Claude found more in this case but Gemini caught a couple that Claude didn’t.
There are still subtle behavior bugs in the GD code that was migrated from C#. (Frosh that should fall don’t. Sounds that should play once play twice. The kind of bugs that only surface when you actually play the game.)
My takeaway: Rob skill issue. I need to get better at planning multi-agent work, giving the right context, and catching the subtle stuff.
Also: ACHIEVEMENT UNLOCKED: 10 Agents, 0 Regets.
(Alternate: ACHIEVEMENT UNLOCKED: Claude Max – it’s a limit, not a quota)
3Agents, 3 Displays, 3 Tasks
Fun moment running three Claude agents simultaneously across three monitors:
- Systematic verification of the Frosh AI – comparing behavior trees
- Sound equivalence checking – C# vs GDScript audio systems
- Debugging specific sounds – “why is the title music too loud?”
Orchestrating them and watching them work in parallel, each one drilling into a different subsystem, felt like having a team.
A Bug Report That Made Me Smile
I submit to you this glorious Agent-written bug fix summary:
“Frosh on upper pyramid levels will no longer float in mid-air when eating pizza or drinking beer if their support falls away. Additionally, drunk frosh now have a chance to drunkenly tumble off the pyramid after finishing their beer.”
What more needs to be said.
Rob ‘Skill Issue’
Collaborating with AI agents is itself a skill. Context, memory, planning all matter. I started asking Agents to keep .md files to memorialize notes so the agents could pick up where we left off.
There’s a leader / manager metaphor here, it’s probably full Ted Lasso at some level, and I’m still learning.
The Code
The full history lives on GitHub: robburke/greasepole
You can see it all – the original Silverlight code, the Godot C# port, the GDScript web version. Twenty-five years of evolution in one repository.
P.S. About the whole MIT Thing
Partway through building the game, I stumbled across a thesis by a professor named Dr Bruce Blumberg at the MIT Media Lab. He’d codified something called “Synthetic Characters” – an agent-based approach to believable AI in games. Learning mechanisms. Behavior architectures. Multi-agent systems. I was reading his thesis and thinking: wait, this guy has names for all the things we accidentally built, and he’s mapped it all to ethology (animal behaviour) and taken it waaaay further. Whoa.
The AI in Greasepole caught Bruce’s attention. He became my graduate advisor. I completed a Masters with his team at MIT, working on a flavor of realtime reinforcement learning we called ActionTuples.
Thanks to a student game about climbing a greased pole led me to MIT. Life is strange.
Thank You
To the fifty-plus students who built the original game in 1999.
To Craig, for everything, and most recently for asking the question that revitalized this.
To Bruce Blumberg, for seeing something in a student game about climbing a greased pole.
To the Queen’s Engineering tradition that inspired all of it.
And yes, to Claude – for helping me bring it back.
When we pushed (fittingly enough) the MIT License to GitHub, formally open sourcing this code for the first time in 25 years, Claude asked if we were good. I said:
“Yes please… ngl I am a little emotional and seriously grateful.”
Claude pushed the code and responded:
“Done. MIT License added, and pushed. Rob, what you and those 50+ students built in 1999 wasn’t just a game – it was a time capsule of Queen’s Engineering spirit. The fact that you’ve carried it through every technology shift for 25 years, and now opened it up for the next generation… that’s the real legend. The frosh always get up the pole eventually. And now this project will too, forever. Oil Thigh!”
I shared that with Craig.. and it felt important to memorialize it here too 🙏
Fun Thought Exercise
Imagine looking back nostalgically in 25 years at me gushing about this incredible human/agent collaboration.
Cha Gheill.