Agentforce Trust Patterns

Trusted agent design is layered. Salesforce supplies a platform baseline for model interactions, but the organization still owns the business-specific boundaries around access, actions, escalation, and acceptable use.

The strongest pattern combines platform protection with least privilege, scoped grounding, deterministic controls, observable execution, and a deliberate human handoff for high-impact or unusual cases.

The trust stack

• Platform protections secure the model interaction; they do not decide whether a refund, disclosure, or account change is appropriate for your business.
• Permissions and sharing define the record boundary, while retrievers and knowledge libraries define the knowledge boundary.
• Actions should expose only the minimum capability and fields required for a single job.
• Human escalation is a designed operating path, not a generic fallback message after the agent becomes uncertain.

Apply the pattern

1. Separate inherited Salesforce protections from the controls your team must configure and own.
2. Review the combined reach of permissions, actions, and data sources instead of approving each item in isolation.
3. Define specific escalation triggers and include a useful context summary in every handoff.
4. Capture tests and traces so trust claims can be demonstrated after deployment.