Start by limiting what an agent can touch. Give it read-only access first, and add write or send permissions only for specific, low-stakes tasks. Work in a copy, a draft folder or a test environment, not in live files.
Keep a human approval step before anything irreversible: sending email, deleting data, spending money, publishing. Ask the agent to explain its plan before it acts, and keep a log of what it did. Check results on a small batch before running the full job. Make sure you can undo changes, with backups or version history. Expect occasional confident errors, and build your process so one error cannot do serious damage.