The agreed fact about the Internet of Things (IoT) is that, within the coming years, billions of resources, such as cars, clothes and foods will be connected to the Internet. However, several challenging issues need to be addressed before the IoT vision becomes a reality. Some open problems are related to the need of building self-organizing and self-adaptive IoT systems. To create IoT applications with these features, this work presents a Framework for Internet of Things (FIoT). Our approach is based on concepts from Multi-Agent Systems (MAS) and Machine Learning Techniques, such as a neural network and evolutionary algorithms. An agent could have characteristics, such as autonomy and social ability, which makes MAS suitable for systems requiring self-organization (SO). Neural networks and algorithms of evolution have been commonly used in robotic studies to provide embodied agents (as robots and sensors) with autonomy and adaptive capabilities. To illustrate the use of FIoT, we derived two different instances from IoT applications: (i) Quantified Things and (ii) Smart Cities. We show how exible points of our framework are instantiated to generate an application.