Actions and targets

An action is declared once in a subagent's actions: block and offered to the LLM as a tool in reasoning.actions (reference/agentforce/agent-script.md). Agent Script only describes the call. The flow, Apex class or prompt template must already be deployed to the org, because publishing doesn't deploy them (reference/dx/deploy-errors.md).

Declaring an action

    actions:
        check_weather:
            description: "Fetch the weather forecast for the resort."
            label: "Check Weather"
            target: "apex://CheckWeather"
            require_user_confirmation: False
            include_in_progress_indicator: True
            progress_indicator_message: "Checking local weather..."
            inputs:
                dateToCheck: object
                    complex_data_type_name: "lightning__dateType"
                    description: "Date to check, yyyy-MM-dd"
                    is_required: True
            outputs:
                maxTemperature: number
                    description: "Maximum temperature in Celsius"
                    is_displayable: True
                    filter_from_agent: False

• descriptions are prompts. The LLM picks tools and fills inputs from them, so say what the action does and when to use it.
• is_displayable lets the output be shown to the user. filter_from_agent: True hides an output from the LLM.
• require_user_confirmation: True makes the agent ask before running a write action.

Target schemes

target: "<scheme>://<API name>". The linter accepts:

Scheme	Calls
flow://	An autolaunched flow, by API name
apex://	An Apex class with an @InvocableMethod
prompt:// (or generatePromptResponse://)	A Prompt Builder template (family 6)
standardInvocableAction://, quickAction://	Built-in invocable actions; quick actions
apexRest://, auraEnabled://, api://, externalService://, externalConnector://	Apex REST methods, @AuraEnabled methods, APIs, External Services
mcpTool://, platformMcpTool://	MCP tools (sf agent mcp manages the API Catalog)
retriever://	A Data 360 retriever. Its inputs aren't documented; use a Data Library instead (reference/data360/search-indexes-and-retrievers.md)
cdpMlPrediction://	A Data 360 / Model Builder prediction
namedQuery://, expressionSet://, runExpressionSet://, decisionTableAction://, integrationProcedureAction://, executeIntegrationProcedure://, serviceCatalog://, createCatalogItemRequest://, slack://	Industry and platform actions
placeholder://	A stub for design; the linter warns

Any other scheme fails, for example: "Action 'verify' uses unsupported target scheme "lambda://". Supported schemes: api, apex, …". The linter doesn't check that the target exists. A wrong API name only fails on sf agent validate, publish or a live preview.

Matching names to the implementation

Target	Input names must match	Outputs
apex://Class	The @InvocableVariable fields of the request class	The fields of the response class
flow://Flow	The flow's input variables (available for input)	Its output variables
prompt://Template	"Input:<apiName>", quoted, one per template input	promptResponse: string

The CLI's agent template pairs this Apex class with the check_weather action above:

public with sharing class CheckWeather {
    @InvocableMethod(label='Check Weather' description='Check weather at the resort for a specific date')
    public static List<WeatherResponse> getWeather(List<WeatherRequest> requests) {
        WeatherResponse response = new WeatherResponse();
        response.maxTemperature = 21.5;
        return new List<WeatherResponse>{ response };
    }
    public class WeatherRequest {
        @InvocableVariable(required=true description='Date to check, yyyy-MM-dd')
        public Date dateToCheck;
    }
    public class WeatherResponse {
        @InvocableVariable(description='Maximum temperature in Celsius')
        public Decimal maxTemperature;
    }
}

The same template calls a prompt template with:

            check_events: @actions.check_events
                available when @variables.guest_interests != ""
                with "Input:Event_Type" = @variables.guest_interests

Types and complex_data_type_name

• Plain types: string, number, boolean, date, datetime, time, currency, id, integer, long, list[...].
• object and list[object] need complex_data_type_name:
• Lightning types: lightning__dateType, lightning__dateTimeType, lightning__integerType, lightning__doubleType, lightning__booleanType, lightning__objectType, lightning__listType.
• An Apex class: "@apexClassType/c__Outer$Inner".
• Without one, the linter warns: "has type 'object' but lacks 'complex_data_type_name'".
• On a primitive it's redundant, and the linter warns: "has primitive type 'string' and does not require 'complex_data_type_name'". Salesforce's own agent template triggers this warning on its prompt template action.
• An Apex Integer output generally needs object with lightning__integerType, per the recipes' rules.

What publishing creates

sf agent publish authoring-bundle compiles the script, creates the agent through the authoring API, then retrieves the generated metadata into your project:

Org API version	Retrieved
68.0 (Winter '27) and later	AiAgentDefinition + AiAgentDefinitionVersion, with the action dependencies (flows, Apex, prompt templates)
Earlier	Bot (+ BotVersion), one GenAiPlugin per subagent, one GenAiFunction per action, and the Agent record

Commit the generated metadata with the bundle, so a deploy to another org includes the agent's Bot/BotVersion (reference/dx/agentforce-metadata-types.md). The full lifecycle: guides/agentforce/build-agent-with-agentforce-dx.md.