Updating AI agent model without breaking things

Back up before making changes. The whole system should be backed up automatically every day, and an important change needs its own separate backup Several models make automatic backups for me at least once a week. Before every major update, I make a manual backup. I update my Claude Code agents manually when a major new model comes out. The last time, I moved them from Opus 5 to Opus 5.5 and skipped the versions in between You decide when to update your agents. You don't have to install every new model as soon as it appears. But you do need a backup before making changes What to save before an update A backup is a copy you can use to restore a working system. Save more than the model settings: include the agent's instructions, memory, code, data and connection settings. Otherwise, switching back to the old model could leave an important part of the system in its new state Check the backup date and try restoring it in a separate environment. Keep copies separate from the main system and protect access to them: they may contain personal data and connection settings. CISA recommends regularly testing recovery and keeping backups separate How to test the new model Run it separately from the working agent first. Give it a few everyday tasks and difficult cases where accuracy matters. If the agent manages a calendar, for example, check whether it understands the date, finds available time and asks for confirmation before creating a meeting Disable real sending and data changes during tests, or use safe test records. Compare its answers and actions with the old model: does it follow instructions, use tools correctly and carry out its checks? Look at speed and cost too - a stronger model on paper isn't always the better fit for your task Check the model provider's instructions: behavior and settings can differ from the previous version. For example, Anthropic's Opus 5.5 guide describes differences from Opus 5 and how to set reasoning effort After switching Move one agent at a time rather than the whole system at once. Check its real tasks and errors before moving to the next one. If something works worse, restore the old model and settings and, if necessary, data from the tested backup. Actions already taken in external services may need to be corrected separately Changing models is part of AI system support and development. What matters is how the model handles your tasks and whether you can return to a working state, not the version number