Start by limiting the actions the agent is allowed to take and the systems it can reach. Use an identifiable account, narrow permissions, and approval steps for sensitive or difficult-to-reverse actions. Keep useful activity records and provide a way to pause the workflow when behavior is unexpected. Assign an owner for configuration changes and incidents. Before broader rollout, test ordinary business cases and failure handling in an appropriate controlled setting. An assistant that drafts a suggestion and an agent that changes customer records create different risks; the access and oversight should reflect that difference.
Related: User access reviews.
Further reading: NIST AI Risk Management Framework; NIST access-control guidance.