DAGGER: Dynamic Agentic Graph-Guided Efficient Reasoner for Container Image Dependency Optimization

Abstract

Container image (CI) bloat degrades cloud deployment efficiency, as structural redundancy and obsolete package buildup over time result in complex dependency graphs that remain largely unaddressed by current CI optimization methods. Traditional visualization techniques rely on user expertise for manual optimization, whereas runtime methods require long execution traces to identify and retain only accessed dependencies. Conversely, static-analysis-based methods lack semantic visibility in dynamic loading or reflection contexts, leading to over- or under-approximations. To address these limitations, we propose Dynamic Agentic Graph-Guided Efficient Reasoner (DAGGER) that dynamically selects from an ensemble of optimization strategies based on CI structures and dependency patterns. DAGGER’s agentic AI module utilizes knowledge retrieval on the graph representation of CI dependencies to analyze and prevent the removal of critical dependencies, ensuring zero-loss operational functionality. DAGGER’s knowledge base is built using curated examples that its agentic AI module uses to generate optimized CI. Based on preliminary experiments on production CIs, DAGGER achieves an average size reduction of 35.7% with a 0.2% breakage rate, compared with state-of-the-art baselines.