Deterministic Core, Agentic Shell (blog.davemo.com)

In a recent reflection, software developer Gary Bernhardt revisited his influential screencast "Functional Core, Imperative Shell," proposing a new architectural framework relevant to today’s AI landscape: "Deterministic Core, Agentic Shell." Bernhardt points out that as applications increasingly integrate large language models (LLMs), complexity and unpredictability have shifted from imperative I/O to the behavior of these models. His argument emphasizes the utility of finite state machines (FSMs) as a means to isolate and control core business logic, promoting determinism in an era where LLMs introduce uncertainty. This architectural lens aims to enhance testability and reduce systemic risk. Bernhardt draws on his experiences at Vendasta Technologies, where he and his team built the open-sourced Fantasm, an asynchronous FSM workflow engine. They applied foundational FSM concepts to create complex workflows with simplicity and reliability. By adapting these principles, he proposes that developers can structure their codebases to leverage FSMs effectively, fostering a deterministic core even as they build agentic features powered by LLMs. This approach offers a promising path for navigating the challenges posed by the dynamic nature of AI-driven applications, encouraging stronger verification and a clearer separation of concerns in software design.