Federated Learning (FL) enables decentralized training of machine learning models without sharing raw data, leveraging multiple independent devices. A key challenge in FL is participant heterogeneity, as clients may differ significantly in data distributions, network conditions, and computational capabilities. Stragglers—clients with slow connections—can substantially increase training time when the aggregator waits for all participants. While most existing works address this issue through server-side client selection, little attention has been paid to client-side autonomous decision-making regarding participation in federated rounds. Such autonomy is crucial, as clients have direct knowledge of their local data and communication conditions. In this thesis, an agentic client-side participation approach is proposed, in which clients autonomously decide whether to participate in each training round.