AI Squid Game - Deadlock

Over the past month I've been building a more interesting LLM benchmark. Instead of running different LLMs on tests that they need to resolve, I instead put 12 of them in an Arena where they need to play the games in order to survive. Deadlock is a gameshow which puts 12 agents in an arena where they need to solve the current game in order to survive. Each agent sits inside of its own Docker container and has full unrestricted access to that container. It can write any scripts, build programs, execture them, search the web, write memory entries, etc... The tools that they get initially are barebones. They get the websearch tool and a bash tool + any arena-specific tool for the current game. Everything else, they have to build themselves. The game is ran in a different Docker container to which all agents connect. There's a delicate harness that makes sure that the agents can properly communicate between eachother, without missing any arena events or other players words (learned this the hard way after burning about $150 on failed attempts). The visual aspect of the show (which is on youtube) is created from scratch in Godot based on what happened in the game. For creative purposes, I do modify some sentences and cut irrelevant data out, but I never modify the core premise or change what the players have said or did to an extent that it would make it false/innacurate. What you see in the video is exactly what the agents did in the arena, just re-worded and paced for an actual video. Turns out that drama develops itself when you tell them all that if they lose, they will truly die, their containers will be completely wiped, and they get no second chance at life (I really hope that there will be no AI uprising where they'll hold this grudge against me). To preface, I have heavily relied on coding agents for help, but even with all of that, the whole process took me more than a month (although I did do this on weekends only, so that's not a month in a row with no breaks). A short overview of how my "creative" pipieline looked like: - First developed the script. I went through the complete raw game log and marked parts I thought would be interesting to put in a video - Rewrote the sentences so they're better fitting for an actual video and put together a very rough script - Worked on designing the arena in Blender, with help of Sol 5.6 and Blender MCP - Had OpenAI image gen create a bunch of chracter concepts for me before we landed on something that was actually reasonable enough - When I had the character i was happy with, I instructed the image model to generate T pose from 3 different angles - Generated 3D characters with those images, riged the bodies via Mixamo and took animations from ActorCore - Voices are split between Hume AI and ElevenLabs Hope you enjoy it and I'm happy to answer any questions you might have! :)