Large-scale distributed systems can be composed of multiple components and computing tiers, all of which have unique contributions to the system's high-level objectives. Consider, for instance, that videos captured by surveillance cameras can be processed on nearby devices and then streamed over the internet. To ensure the system's functionality, stakeholders describe each tier's expected behavior through Service Level Objectives (SLOs), e.g., maintaining processing latency under a certain boundary. Evaluating these SLOs requires a set of metrics (i.e., sensory observations), which are commonly collected at one central system location; given this data, it can then be calculated to what degree SLOs were fulfilled. However, this requires transferring massive amounts of data; further, the latency for detecting and resolving SLO violations is high. The human body, as an example of a complex system, would collapse from the overhead of evaluating each cell's requirements centrally (i.e., in the brain); hence, requirements assurance must be decentralized to the respective system components. Given that its SLOs are violated, intelligent components should detect and resolve this autonomously by interacting with their environment. Understanding how to adapt a system requires in-depth knowledge, which can be provided through Machine Learning (ML) techniques. In particular, this can involve extracting causal relationships between components. However, training such models requires large amounts of data upfront, which are also subject to variable shifts over time. This promotes the usage of Active Inference (AIF) for two reasons: (1) AIF provides continuous model accuracy while creating world models without prior assumptions, and (2) AIF agents aim to persist over time, thus they can be used to modify the system according to expected SLO fulfillment and model improvement.