9.3 Agent Builder, Prompt Builder, and Grounding Concepts

Builder concepts administrators should know

Agentforce Builder is the workspace for creating and customizing agents. Current Salesforce guidance also describes Agentforce Studio as a central app for building, testing, and monitoring agents, with access from the App Launcher when the org and user have the required feature access. Some older or transition documentation may still use the term topics where newer material uses subagents. For admin study, focus on the concept: a defined area of work that gives the agent instructions and allowed actions.

An agent configuration usually includes agent-level instructions, conversation behavior, subagents or topics, actions, variables, grounding sources, preview testing, and deployment settings. The admin should understand each part without needing to write Apex. Agent-level instructions define the agent identity, tone, boundaries, and refusal rules. Subagents or topics help route the user request to the right skill area. Actions let the agent do work, such as retrieve data, call a flow, create a record, or hand off a conversation.

Prompt Builder is related but not identical. Prompt Builder helps create prompt templates that can use Salesforce record fields, related data, and instructions to generate content in a repeatable way. For example, a sales team might use a prompt template to draft a follow-up email from opportunity details, recent activities, and approved messaging. An agent may use prompts as part of a broader conversational experience, but the admin should still separate template design from agent orchestration.

Grounding is the admin's bridge between generative output and trusted business context. A generic model can produce fluent text, but a business agent needs approved facts. Grounding can connect the agent to CRM records, knowledge articles, unstructured documents, or data libraries depending on the feature set. The admin should ask who owns each source, whether it is current, whether it is approved for the audience, and whether permission rules are enforced at runtime.

A strong grounding strategy is selective. Do not connect every document because more data feels safer. Too much uncurated content can make answers less reliable and harder to govern. Start with published knowledge, clear policy documents, or records with strong data quality. Remove duplicates, stale drafts, obsolete pricing, old procedure files, and confidential notes that do not belong in the agent's context.

Builder workflow for an admin-led pilot:

1. Confirm feature enablement, licensing, app visibility, and builder permissions.
2. Create or choose the right agent type or template for the use case.
3. Write agent-level instructions that define scope, tone, refusal behavior, and escalation.
4. Add subagents or topics for each focused business area.
5. Add grounded data sources and verify source ownership, publication status, and access rules.
6. Add actions only after reviewing validation rules, flow context, user permissions, and side effects.
7. Preview with happy paths, incomplete requests, restricted data requests, and unsupported tasks.
8. Save versions and document what changed before deployment.

Developer handoff is part of admin judgment. Flow actions may be admin-configurable when the process already exists and is safe to expose. Apex actions, external service calls, complex API integrations, custom data retrieval, and advanced error handling may require a developer or architect. The admin is still responsible for asking the right questions: what user context is used, what data leaves Salesforce, how errors are surfaced, and how the action is tested.

Prompt design has governance rules too. A good prompt template includes role, goal, data inputs, output format, and constraints. It should tell the AI what to do when information is missing, not encourage guessing. If the template drafts customer-facing content, the admin should include review steps, approved language, and brand or compliance constraints. For sensitive messages, AI can draft, but a human may still need to send.

Study trap: do not treat grounding as the same thing as permission. Grounding selects context sources, while permissions decide who can access records, fields, and actions. Both must be correct. Also do not assume a prompt template is safe because it works for an admin. Test with the users and records the business actually intends to support.