AI Agent Builds Doom-Playing Network Modeled on Fruit Fly Brain Connectome

A developer used AWS's open-source Strands coding agent to build a neural network constrained to the structural wiring of a real fruit fly brain, comprising 49,393 neurons and 9 million synapses, and trained it to play the video game Doom. The network is not a biological fly but mirrors its connectome, with standard convolutional and MLP layers handling more than half the parameters. The entire project — roughly 3,300 lines of Python and shell code, 28 commits, and multiple training runs — was generated by the Strands agent over five days with minimal human input. The developer's prompts across seven sessions were largely limited to brief check-ins and approvals, while the agent independently managed training pipelines, cloud GPU deployments across three AWS regions, and version control. The project highlights how AI coding agents can drive complex machine-learning infrastructure end-to-end from a single initial prompt.