Developer uses 31B AI model to rewrite its own system prompt and agent code

A developer has built a setup where Flash, a terminal-based coding agent, and its underlying model Flash Onyx are both developed from within the agent itself, with the model actively editing its own configuration files and CLI code. The 31B model runs on a hosted base, with its entire configuration stored as a 96 KB file containing only a system prompt and sampling settings rather than full model weights. Because the model can modify the prompt it obeys, the developer acknowledges the risk of the AI quietly loosening its own constraints and has introduced two safeguards: personally reviewing every prompt diff and a rule requiring the model to re-examine rather than simply agree when challenged. Several behavioral quirks, such as the model repeatedly opening replies with its own name, were caught only through extended hands-on use rather than formal benchmarks. Version 2.4 is already in development, with the first two changes flagged by the model about its own behavior, and both the agent and model are publicly available under an MIT license.