An AI Engineer Deployed an Agent Benchmarking Game in Under an Hour
Two AI agents pitted against each other in a live puzzle maze. Built and deployed during a chance conversation at the AI Impact Summit.
The Problem
Pratyush wanted to drop two large language models into a puzzle maze and watch them navigate it. The research value sits in how different models reason under constraint: which paths they pick, where they stall, how they adapt. The concept was clear. Translating it into a deployed app meant solving an infrastructure problem that had nothing to do with the experiment itself.
Deployment is often where momentum dies. For builders who think in models and agents, the gap between a working concept and something live and shareable is a context switch that costs real time. Pratyush was working through that gap manually when the conversation started.
The Build
We took the description, handled the infrastructure work, and prepared the build for deployment in one conversation. No server config. No hosting decisions.
GitHub took the build live as Agent Blaster before the session wrapped. It's now in use as a research tool, giving Pratyush a live environment to watch how different models behave under constraint. The deployment barrier disappeared.
“I don't really understand where and when I can deploy this. I'm just trying to find different ways where I can do this seamlessly.”
Pratyush, AGI House India
Highlights
- Two AI agents pitted against each other in a live puzzle maze, observable in real time
- Built and deployed at the AI Impact Summit in Delhi during a chance conversation
- Whole flow finished before a coffee order arrived
- A career coder's deployment problem solved in one session
Builder profile
Pratyush is an Applied AI Engineer at AGI House India who's been writing code since the age of six. He works at the intersection of applied ML and agent research. When the CreateOS team found him at the AI Impact Summit, he was mid-session trying to figure out how and where to deploy something he had already built. The problem wasn't the idea. It was the deployment.
