The K-alibra Agent is a Language Model (LM)-based orchestrator designed to dynamically adjust the number of participating clients in each round of Federated Learning (FL). It was created to overcome the rigidity of traditional static algorithms, which typically use a fixed number of clients, a practice that can lead to inefficiency or network overload.
The project was developed through a partnership between researchers from the State University of Campinas (UNICAMP), the Federal University of Pará (UFPA), and the Federal University of Minas Gerais (UFMG), all of which are part of ICoNIoT. They are: Rafael O. Jarczewski (UNICAMP), Eduardo Cerqueira (UFPA), Antonio Loureiro (UFMG), Leandro A. Villas (UNICAMP), and Allan de Souza (UNICAMP). It is published in the proceedings of SBRC 2026 and can be read in full here.
Federated Learning (FL) is a way to train artificial intelligence securely, since each person’s data remains protected on their own devices, without needing to be sent to a central server. The problem is that this process tends to use a lot of data and becomes difficult to manage when there are many users.
Currently, to address this issue, systems select which devices will participate in the training. However, these systems are “stubborn”: they tend to always use the same number of devices, without adjusting that number based on current needs. This causes the process to waste resources or take longer to learn.
To overcome this rigidity, K-Agent was created. It functions as an intelligent “boss” (using language models similar to those behind AI chatbots) that dynamically decides how many devices should participate in each stage.
It works in three steps:
- It assesses the current state of the training.
- It reasons to find the best strategy.
- It acts by determining the optimal number of participants.
Tests have shown that K-Agent is highly efficient: it can reduce internet usage by between 44.4% and 59% compared to traditional methods, while maintaining stable, high-quality learning. Additionally, it can explain the reasoning behind its decisions, making the system more transparent for developers.
The project has been published in the proceedings of SBRC 2026 and can be read in full here.