Agentforce security designed before the first action.

Define what the agent must not do first.

Every agent needs a clear goal, allowed topics, approved knowledge and a constrained action set. Informational responses should be separated from operations that change data. Greater autonomy is appropriate only when actions are low risk and reversible.

Financial commitments, access changes, regulated decisions and customer promises need stronger controls. We design confirmations, approvals, limits, escalation and a safe stop path.

• Least privilege and sharing
• Allowed topics and actions
• Human approval and escalation
• Validation, limits and idempotency

The Trust Layer does not replace correct Salesforce configuration.

The Einstein Trust Layer provides mechanisms including secure grounding, sensitive-data masking, audit and supported zero-data-retention arrangements. It cannot repair overbroad permission sets, an incorrect sharing model or an integration that bypasses access control.

We review profiles, permission sets, sharing, Named Credentials, integration identities and document access. We also verify whose context an agent acts in and what different roles can see from the same interaction.

Production needs an owner, metrics and a controlled change path.

Governance defines who approves a new topic, action or knowledge source, which tests must pass and who responds to incidents. It combines Salesforce release management with evaluation of non-deterministic agent behaviour.

After launch, teams monitor classification failures, action errors, escalation, answer quality, consumption and security signals. Instructions and knowledge changes should be versioned, tested and reversible.