The advent of complex and physically distributed systems and the need to minimize the down-time of services and production processes call for more efficient supervision systems. Traditionally centralized, the anytime supervision of such systems is challenged when communications between the supervision and the supervised systems become either slow, disrupted or too costly. In this paper, we propose an agent decision model that allows a multi-agent supervision system to dynamically adapt itself to the state of the communications by means of distributing the diagnosis an repair process. Experiments on a simulator for distributed systems using an industrial dataset show that our proposal does lead to an anytime and adaptive supervision of distributed systems where a short response time prevails over a limited repair extra-cost.