Agentforce Fundamentals and Capabilities
What Agentforce is
Agentforce is Salesforce's platform for building and running AI agents that reason over data, choose a plan, and execute actions through Salesforce metadata. The Atlas Reasoning Engine orchestrates the loop: interpret the request, select a topic, retrieve grounding, choose actions, and respond. The LLM supplies language understanding and generation; Atlas supplies planning and tool use. Agents are configured in Agent Builder with topics, instructions, actions, channels, and guardrails, and they run under the Einstein Trust Layer.
Concepts to know
- Topics define a job-to-be-done and contain a classification description, scope, instructions, and actions.
- Actions are the “how”: Flow, Apex, prompt template, or standard action with typed inputs and outputs.
- Channels include messaging, web, mobile, and service consoles.
- Human handoff and escalation are first-class capabilities.
Traps
Questions often blur topic and action, or assume the LLM executes code. It does not: actions are deterministic and permissioned. Another trap is treating Agentforce as a chatbot only; agents can be autonomous and multi-step. Also, Atlas is not a model you pick; it is the reasoning layer.
Worked example
“Where is my order?” → classification selects the Order Status topic → Atlas calls a Flow action with an order ID input → the Flow queries the Order object → the response is grounded and permission-aware. Rule of thumb: topic = what the agent is allowed to do; action = how it does it.