Agent Builder is the no-code interface for building Agentforce AI agents — defining their topics (when the agent should engage), actions (what the agent can do — call Apex, query records, run flows), guardrails (what the agent must NOT do), and channels (where the agent appears). Agentforce is Salesforce's autonomous AI platform; Agent Builder is the visual tool to configure agents without writing prompts directly.
- Confirm Agentforce is licensed
Agentforce is a paid product line. Setup → Agents — if absent, Agentforce isn't enabled in your org.
- Open Setup → Agents → Agent Builder (or App Launcher → Agentforce)
Navigation may vary by Salesforce release. Look for Agentforce or Agents under Setup.
- Click New Agent
Wizard walks through agent type — Customer Agent / Sales Agent / Service Agent / Custom.
- Set Agent Name, Description, and Welcome Message
Welcome is what users / customers see when the agent first engages.
- Configure Topics
Topics are the situations the agent should handle — "Order Status", "Password Reset", "Product Recommendation." For each topic, specify when the agent invokes.
- Configure Actions per Topic
Actions are what the agent can DO — call an Apex method, query a record, run a Flow, post a Chatter message. Each action is bound to a topic.
- Set Guardrails
Constraints on what the agent must NOT do — "never quote pricing without checking current Price Book," "never share Account-level data with anonymous users." Guardrails are the safety rails.
- Pick Channels
Where the agent appears: Embedded Service / Slack / WhatsApp / Salesforce Mobile / etc.
- Test in the Agent Builder Test Pane
Before activating, simulate user inputs and watch the agent's reasoning + actions. Iterate.
- Activate
Live to users in the configured channels. Monitor Setup → Agent Performance for engagement metrics.
Customer / Sales / Service / Custom. Locked at creation.
When the agent engages.
What the agent can do (Apex / SOQL / Flow / Chatter / etc.).
Behavioral constraints.
Embedded Service / Slack / WhatsApp / etc.
- Agentforce is a paid product line on top of Service / Sales Cloud. Confirm licensing before scoping — many features overlap with Einstein but with different licensing.
- Guardrails are critical and easy to under-specify. An agent without good guardrails can hallucinate, share wrong data, or take unauthorized actions. Spend time stress-testing guardrails before going live.
- Agent reasoning is non-deterministic. Two users asking the same question can get different responses depending on conversation context. Test with realistic conversation paths, not just single questions.