This paper presents an architecture and implementation of a self-healing sandbox for the execution of third party code dynamically loaded which may potentially put in risk application stability. By executing code in a fault contained sandbox, no faults are propagated to the trusted part of the application. The sandbox is monitored by a control loop that is able to predict and avoid known types of faults. If the sandbox crashes or hangs, it can be automatically recovered to normal activity without needing to stop the main application. A comparison between an implementation of the sandbox in a domain-based isolation and operating-system based isolation analyses performance overhead, memory footprint and sandbox reboot time in both approaches. The implementation has been tested in a simulation of an RFID and sensor-based application.