When a job with the SSHAgentBuildWrapper enabled fails very early (for instance during SCM checkout), an ssh-agent process is left behind. The issue is that the SSHAgentEnvironment is instantiated very early (from preCheckout), but its tearDown method will only be called if execution reaches BuildExecution.doRun (which comes after the SCM checkout phase in AbstractBuildExecution.run).
Before ssh-agent-plugin 1.14, there was no ssh-agent process, so the issue with some SSHAgentEnvironment not being teared down was less visible (but probably there was already some other kind of less obvious resources leaks with AgentServer not being properly closed).
This kind of issue with some Environment not being properly teared down can happen as soon as they are not instantiated from BuildWrapper.setUp, but from earlier phases (like BuildWrapper.preCheckout or RunListener.setUpEnvironment). As such, maybe that's something that should be fixed in core (maybe in AbstractBuildExecution.run) rather than specifically in the ssh-agent-plugin, I don't know...
I've written and attached a "generic workaround" RunListener, which tries to detect this situation from onComplete, and call tearDown for all Environment if it has not been done already. It's not something I propose for inclusion, but rather some code to exhibit the issue. If an ssh-agent specific fix is desirable, then a similar approach might be an option (but targeting SSHAgentEnvironment only).