Building a multiplayer game world with autonomous Fable-class models (simonwillison.net)

Jacob Sansbury's recent blog post unveils the ambitious project Spawn, a massively multiplayer gaming environment that utilizes autonomous Fable-class AI agents, dubbed "Savi." Players can interactively modify the game world by issuing prompts, such as creating castles or summoning dragons. What sets Spawn apart is its innovative use of a swarm of AI agents that autonomously manage live code, implementing changes and resolving bugs in real-time, reportedly at a cost of over $3 million in tokens monthly. This continuous deployment approach enables rapid iteration and deployment of features, where changes are delivered to users within an hour, with an impressive rollback capability of just 30 seconds if issues arise. The significance of this development lies in its pioneering application of agentic engineering and advanced AI models to enhance game design and functionality. Jacob highlights the use of Fable 5 as crucial for achieving such autonomy, while his experimentation with other AI models, such as GPT-6 Astra, showcased the limitations and challenges when security measures interfere with functionality. This case illustrates the potential for using gaming environments as experimental grounds for exploring the frontiers of AI-driven software development, raising intriguing questions about the future of coding practices and the role of AI in more critical software domains.