AI agents make mistakes. Here's the production architecture insurance agencies use to catch them — approval gates, dry-run mode, audit trails.
Any AI deployment plan that doesn't account for "what happens when it's wrong" is incomplete. Here's the production-grade pattern.
1. Approval Gates on Writes
Read operations: free-flowing. Write operations: queued for human review. Claude drafts the SMS; CSR taps approve. Claude proposes the task; producer accepts. The agent does the work; the human owns the decision.
2. Dry-Run Mode
Before any bulk operation, the agent describes what it's about to do — affected record count, fields changing, sample diff. Human confirms or aborts. unLocked exposes a dry_run: true parameter on bulk ops for exactly this.
3. Audit Trail Per Action
Every CRM write tags the agent ID, the human who approved (if applicable), and the prompt that triggered it. Searchable in the activity timeline. Required for any post-incident review.
4. Rollback Patterns
For destructive operations (delete, mass-update), the API returns a rollback token. Within 24 hours, a single API call reverses the change. Used sparingly but invaluable when an agent goes off-script.
5. Rate Limits Per Key
Each API key has configurable rate limits. An agent that suddenly tries to update 10,000 records in a minute hits a soft cap and pages the admin. Prevents runaway loops.
6. Sandbox Environments
unLocked offers separate sandbox accounts. Test new agents against sandbox data before pointing them at production.