Agentic federated learning for dynamic privacy budget management

Federated learning enables collaborative model training without centralizing raw data, supporting privacy-sensitive applications ranging from mobile assistants to healthcare systems. However, model updates may still leak sensitive information. Differential privacy mitigates this risk by injecting noise into local updates, although existing approaches typically rely on static, globally scheduled, or fixed client-level privacy budgets that may poorly align privacy expenditure with learning utility. This work introduces an agentic privacy management framework in which autonomous client-side agents dynamically allocate privacy budgets based on gradient informativeness, loss variation, and remaining budget constraints. Unlike predefined allocation schemes, the proposed mechanism treats privacy spending as a state-aware and context-aware control process embedded within federated learning. Experimental results show consistent improvements in convergence and final accuracy compared to representative differential privacy allocation baselines, while reducing cumulative privacy expenditure.