Towards the Versioning of LLM-Agent-Based Software

Large Language Models (LLMs) are redefining software practices by introducing agents that adapt rapidly to new data, tasks, and contexts. Conventional version control, optimized for source code updates, struggles to capture the evolving states of LLMs, prompts, training data, and agent behaviors. This paper proposes a multi-artifact approach that marries traditional code repositories with specialized registries for model weights, datasets, and prompt templates, tied together by metadata-rich descriptors. Through analogies to familiar computing layers and real-world scenarios, we illuminate the challenges of non-deterministic outputs, layered dependencies, and shifting user expectations. Our framework supports both "strict modes" for reproducibility in high-stakes contexts and "fluid modes" allowing iterative improvements with lower risk. By unifying versioning policies and tools across technical, organizational, and user-centric dimensions, we aim to foster more robust, transparent, and ethically informed LLM-agent systems that can reliably evolve alongside their expanding capabilities.