Memory hook: The agent can only be useful if its data and actions are trustworthy.
Must remember
Agentforce supports AI agents that interpret requests, use configured subagents/instructions (called topics in older material) and invoke allowed actions. Good use cases have a clear outcome, reliable accessible data, bounded actions and a human fallback. A deterministic calculation or approval rule may be better handled by ordinary automation than an open-ended language-model decision.
Grounding supplies relevant trusted business context. Poor or inaccessible records can produce weak answers even with a well-written prompt. Separate an agent's instructions from data it reads; customer text or documents can contain misleading instructions and must not override authorized behavior.
Actions can depend on Flow, Apex or other supported integrations, with their own permissions and execution context. Troubleshoot the complete chain: user/agent identity, license/feature access, subagent/action availability, object/field/record access, flow permission/context and external connection. Broad administrator access may hide the original defect and create unnecessary risk.
Agent Builder supports maintaining instructions/prompts and testing conversations in the available experience. Preview ordinary, ambiguous, unauthorized and failure scenarios in a sandbox: testing can execute actions and consume usage credits. Check the actual records/actions used and the resulting side effects, not only the wording of the response. A plausible sentence saying an update succeeded is not proof that it did.
Version and test changes before deployment, including installed reusable prompts/actions from trusted sources. Define handoff and approval for consequential operations. Monitor quality, incorrect responses, denied actions and unexpected data access. AI suitability depends on business risk and evidence, not on whether a feature can be enabled with a toggle.
Choose under exam pressure
| Requirement | Choice and reason |
|---|---|
| Fixed field calculation with exact behavior | Formula/Flow rather than unnecessary generative reasoning. |
| Agent cannot update a record | Inspect its identity and the entire action/data permission chain. |
| Change agent instructions | Version, preview representative conversations and verify actual outcomes. |
Traps
- A good prompt cannot grant missing data access.
- A confident response is not evidence that an action completed.
Active recall
1. What is grounding for?
Providing relevant trusted context for the response or decision.
2. Why bound agent actions?
To limit the effects of incorrect interpretation or untrusted input.
3. First permission-debugging principle?
Trace the actual executing identity through every data and action layer.
4. What should conversation testing include?
Normal, ambiguous, unauthorized and failure cases with side-effect verification.
5. When is a human handoff useful?
When confidence, authorization or business risk exceeds the agent’s permitted scope.