Upgrading dependencies in software projects is a complex multi-step process that requires developers to assess the upgrade decision, understand the changes and their impact, and implement the upgrade and validate it. Although numerous automated approaches have been proposed to support developers, several challenges remain underexplored. Recently, Large Language Models (LLMs) and Agentic AI have pushed the boundaries of automation even further. However, even cutting-edge models have significant limitations, such as hallucinations and context constraints, which increase complexity when applied to real-world systems. These challenges highlight the need for further research to understand both the opportunities and challenges of the AI-based development paradigm. Framed within the scope of a research project (anonymized for review), this vision paper presents a human-centric agentic architecture for end-to-end dependency upgrades where specialized LLM-based agents support different upgrade tasks. We also outline an initial evaluation plan to assess the proposed architecture across dimensions beyond the overall accuracy of the upgrade process. Although the approach warrants further investigation, it is a first step toward designing, validating, and deploying agentic architectures for this use case in real industrial contexts.