Agent Action
An Agent Action is a specific, discrete task that an Agentforce AI agent can perform when interacting with users or processing requests.
Definition
An Agent Action is a specific, discrete task that an Agentforce AI agent can perform when interacting with users or processing requests. Each action is defined with a description, inputs, and outputs, and is mapped to an underlying capability such as a Flow, Apex class, or prompt template. Agent Actions are the building blocks of an AI agent's functionality.
In plain English
“Let's break this down in a fun way: You know how Netflix suggests movies you might like based on what you've watched before? Agent Action does something similar in Salesforce - it uses smart computer programs to look at data and suggest the best thing to do next, saving people a ton of time and guesswork.”
Worked example
A developer at CloudServe defines three Agent Actions for their customer-facing Agentforce agent: "Look Up Order Status" (which calls a Flow to query order records), "Process Return Request" (which invokes an Apex class to create a return case), and "Recommend Product" (which uses a prompt template with Einstein). When a customer asks about their order, the agent selects and executes the appropriate action automatically.
Why Agent Actions are the discrete capabilities AI agents stitch together to accomplish work
An Agentforce AI agent doesn't run on raw freeform language; it runs on a defined catalog of Agent Actions - discrete, named tasks each with declared inputs and outputs. Looking up an Account by name, creating a Case from a conversation, sending a follow-up email, updating an Opportunity stage - each is one Agent Action, mapped to an underlying capability (a Flow, an Apex class, a prompt template). The agent's intelligence comes from selecting and chaining the right actions; the actions themselves are deterministic.
The reason this design pays off is auditability and control. With agents that act on raw natural language, there's no clear answer to what the agent can or can't do; with action-bounded agents, the answer is the action catalog. Define the actions deliberately, document their effects, monitor their usage in production, and the agent's behavior becomes shapeable rather than mysterious. New capabilities are added by defining new actions; risky behaviors are prevented by not defining them.
How organizations use Agent Action
Built five Agent Actions that the support Agentforce agent uses to update customer records - check warranty status, schedule service appointment, file insurance claim, send technical Knowledge article, escalate to human. Each action is one Apex method exposed to the agent with strict input validation.
Use Agent Actions in their reservations Agentforce agent to update Booking__c records and trigger downstream automations. The actions encapsulate the multi-step flow as a single agent-callable function, keeping the agent's prompt tractable.
Audit Agent Actions for permission scope - every action's underlying Apex runs as the agent's user, and the team's policy is that an action only does what the user already could. The audit caught two early-build actions that bypassed user-level field-level security; both got fixed before the agent went live.
Trust & references
Straight from the source - Salesforce's reference material on Agent Action.
- Standard Agent Action ReferenceSalesforce Help
- Best Practices for Agent Action InstructionsSalesforce Help
Hands-on resources to go deeper on Agent Action.
Test your knowledge
Q1. How does the Einstein Trust Layer relate to Agent Action?
Q2. What does Agent Action need to work effectively?
Q3. What technology powers Agent Action in Salesforce?
Discussion
Loading discussion…