Agentforce Fundamentals and the Agentic Enterprise
From assistant to agent
An assistant suggests; an agent acts. In Agentforce, the reasoning layer interprets a constituent's request in natural language, picks a topic, selects an action, fills that action's inputs, executes it, and then explains the result to the user. Almost every question in this domain maps back to that loop.
The building blocks
- Agent type and role - employee-facing agents (caseworkers, service reps) versus constituent-facing agents on messaging, web, or Experience Cloud.
- Topics - the scope of work. A topic carries a name, a description used for routing, instructions that shape behavior inside the scope, and a set of assigned actions.
- Actions - the verbs: autolaunched Flows, Apex, prompt templates, and standard actions such as answering from knowledge or escalating to a human.
- Grounding - retrieving trusted data (Knowledge, Data 360, structured records) at runtime rather than relying on model memory.
- Channels and guardrails - where the agent is deployed and what boundaries constrain it.
Traps worth memorising
First, topics and actions are not interchangeable: topics classify intent, actions do work. Second, the model does not "know" agency policy - ungrounded generative answers are precisely what grounding and the Einstein Trust Layer exist to prevent. Third, Agent Builder (the canvas composing topics and actions) is not Prompt Builder (the tool composing a reusable prompt template that one action may call). Fourth, "agentic" does not mean unsupervised: escalation and confirmation steps are design decisions, not failures.
Rule of thumb
When an agent misbehaves, debug in routing order: topic description → topic selection → action selection → action inputs → action output. Fix only the earliest broken link, then retest.