We've had reliability issues with the EC2 agents. We've not been able to spend the time to isolate the root of the issue. It could be the EC2 plugin. It could be a connection issue between the EC2 agents and the Jenkins master that is running on Microsoft Azure. See INFRA-2548 for further discussion.
In the interim, we rely on users to start their failed jobs again by either closing and reopening the pull request or submitting a new change to the branch.