I worked on a similar issue for a PRPC 6.1 SP2 system about an year back. There is a hot fix (HFix 7083) which addresses this issue. From what I remember this hot fix addresses the issue only for system agents, with the definition of "system agent" being:
Activity name starts with "System" and activity class starts with "Code-" OR Standard agent with Auto Queue management
By this definition, SendCorr is considered as a system agent and the hot fix should handle it.
To handle non-system agents, you can write a custom "system" agent which runs periodically and checks all agents on the node and starts any agent(s) which failed because of an exception.