Agent Topics and Agent Actions: How the Agent Chooses Behavior and Tasks

In Agentforce, an agent may appear to be a single intelligent brain, but under the hood it operates like a governed system. It must decide what kind of work it is doing, which tools it is allowed to use, and which rules apply in the current situation. Agentforce achieves this by separating configuration into two layers: topics and actions. When topics and actions are designed correctly, the agent behaves predictably and safely. When they are not, the agent behaves like a generic language model that tries to be helpful without clear boundaries.

An action is a single, discrete task the agent can perform. It is the operational unit of work. You can think of an action as one atomic step the agent can execute in response to a request. Actions represent real business operations, such as looking up an order, checking delivery status, creating a support case, routing the conversation to a live agent, or updating information when allowed. The key point is that an action is about execution, not about phrasing. It defines what the agent does, not how it talks.

A topic is the category of work that provides structure around actions. Topics do not exist just to label conversations. A topic defines a working context and establishes rules for behavior. Each topic contains actions, plus additional elements that shape how the agent should interpret requests and respond. A topic includes a classification description that helps determine whether a user request belongs to that topic. It includes scope, which defines what is included and excluded so the agent does not overreach. And it includes instructions that define how the agent should behave while operating in that topic. In this way, a topic is a container that bundles allowed actions with the behavioral framework needed to use them correctly.

This is why topics create boundaries, context, and behavioral rules. They set the limits of what the agent should attempt, which reduces confusion and improves reliability. They provide a clear operating frame so even short or ambiguous user messages can be interpreted consistently. And they enforce consistent behavior, because instructions and scope are applied inside the topic context rather than being left to open-ended improvisation.

A critical rule in Agentforce is that the agent can use only the actions assigned to it through topics. Actions do not become available simply because they exist somewhere in the system. They become available when they are placed inside a topic and that topic is connected to the agent. This is the core of action assignment. It is also important to know that a single action can be included in multiple topics. This is a practical design pattern. For example, an action like Create Support Case might be relevant both in an “Order Issues” topic and in a “General Support” topic. Reuse is allowed, but each topic can still enforce its own scope and instructions, which means the same action can be used under different behavioral rules depending on the context.

When the agent “chooses behavior,” it is effectively choosing the correct topic and then selecting the right action within that topic. The reasoning engine evaluates the user request and determines which topic best matches, guided by classification description and available context. Once a topic is selected, the agent sees a constrained set of actions that belong to that topic. It then chooses an action and executes it. If additional steps are needed, the agent continues within the same topic boundaries or switches topics when appropriate. Topics define the mode of operation. Actions define the specific operations that can be executed.

Managing topics and actions happens through the Topics panel in Agentforce Builder. This is where you add and remove topics, place actions inside topics, and control the instructions and scope that govern behavior. There is also an important operational rule: the agent must be deactivated before making these changes. This protects the integrity of the running configuration and ensures updates are applied safely and consistently. In exam scenarios, this detail often matters because it tests whether you understand governance and lifecycle discipline, not only concepts.

In summary, actions are the individual tasks an agent can perform. Topics are the structured categories of work that contain actions and define the rules for how those actions should be chosen and used through classification description, scope, and instructions. The agent can only use actions that have been assigned to it through topics, and the same action can appear in multiple topics. Topics and actions are managed in the Topics panel, and the agent must be deactivated before changes are made.

Next, we move deeper into actions themselves. We will compare standard actions and custom actions, and then explain how to create a custom action using a reference action type, which is a key skill for building enterprise-grade Agentforce agents.