At work, we are 4 sysadmins (two dedicated, two time-shared with other tasks) working in a number of systems. To mitigate this particular problem, we've setup an infrastructure comprised of Kerberos, LDAP and AFS with sudo for privilege escalation.
The intention is to manage the risk of unintentional mistakes causing problems, as well as some degree of traceability when problems occur. No single admin knows all systems, so one may by mistake cause conflicting configs in unknown systems, but in those cases, you can pretty quickly determine who did it, and roughly what was did by examining logs. Enables the shame-factor, education, and acceptably quick recovery.
The highly critical functions runs in an isolated system, upgrades to both code and config are version controlled (as is the bulk of the config of every individual service), further improving traceability, with the praxis to only do updates to those systems in with at least two admins present in order to avoid mistakes and ensure everyone keeps well familiar with those systems.
However, it's very important to note that all of this is intended to improve maintainability, NOT security. All four admins have individual physical access, so any one of us could naturally while doing onsite jobs, perform an "upgrade-reboot", and sneak in init=/bin/bash on any system and do whatever without logging. In the end, you pretty much have to trust your co-workers, or pay in both money and flexibility for a non-centralized organisation.
