Agentforce Notes

Architect Notes
  • https://architect.salesforce.com/docs/architect/decision-guides/guide/determining-agentic-vs-traditional-workflow-automation.html?d=afx
Agentforce Data Library
Conceptual Notes
  • Prompt builder can be used in apex like this:
ConnectApi.EinsteinPromptTemplateGenerationsInput input = new ConnectApi.EinsteinPromptTemplateGenerationsInput();
            input.additionalConfig = new ConnectApi.EinsteinLlmAdditionalConfigInput();
            input.isPreview = false;

            Map<String, ConnectApi.WrappedValue> inputParams =
                new Map<String, ConnectApi.WrappedValue>();

            ConnectApi.WrappedValue oppVal = new ConnectApi.WrappedValue();
            oppVal.value = new Map<String, String>{
                'id' => String.valueOf(RecordId)
            };
            inputParams.put('Input:Opp', oppVal);

            ConnectApi.WrappedValue searchVal = new ConnectApi.WrappedValue();
            searchVal.value = searchQuery;
            inputParams.put('Input:SearchQuery', searchVal);

            ConnectApi.WrappedValue retrieverVal = new ConnectApi.WrappedValue();
            retrieverVal.value = DEFAULT_DATA_LIBRARY_RETRIEVER;
            inputParams.put('Input:RetrieverIdOrName', retrieverVal);

            input.inputParams = inputParams;

            ConnectApi.EinsteinPromptTemplateGenerationsRepresentation result =
                ConnectApi.EinsteinLLM.generateMessagesForPromptTemplate(
                    promptTemplateDeveloperName.trim(),
                    input
                );

            ConnectApi.EinsteinLLMGenerationItemOutput generation = result.generations[0];
            return parseRequirementsJson(generation.text);
  • If we create an agentforce agent with type Employee then we can show in global action on top. You can add ADL in data in the agent

Instruction Canvas:

    • Use / to put IF/ELSE

    • Use @ to use varible and actions

Instruction Writing Rules (Golden Rules)

Rule Why
Write instructions in imperative form “Always call X” not “X should be called”
Keep each instruction to one idea Multi-idea instructions confuse the LLM
Be explicit about order “First do X, then do Y, then respond”
Define fallback behaviour “If the action returns no data, tell the user you couldn’t find the record”
Limit scope creep Add “Only handle X” to prevent the topic from expanding
  • Set negative scope in instructions

How Atlas Actually Selects a Topic:

User message comes in
↓
Atlas reads ALL topic classification descriptions at once
↓
Scores each topic for semantic similarity to the message
↓
Picks the HIGHEST scoring topic
↓
Executes that topic’s actions + instructions

The classification description VS Topic Instruction:-

  • Topic description talks to Atlas (the AI router). Instructions talk to the agent (after routing).
  • The classification description is never seen during execution. The instructions are never seen during routing.

How Atlas works with Apex Action

The @InvocableMethod Description Does the Heavy Lifting

Remember the description you wrote on the Apex action:

apex

@InvocableMethod(
    label='Create Support Case'
    description='Creates a new Salesforce support case for 
    the customer. Call this only after collecting the issue 
    description, affected product, and severity from the 
    customer and after they have confirmed the details.'
)

What a Salesforc Developer need to learn actually:

Skill Your status
Apex + @InvocableMethod ✅ Already know it
Salesforce data model ✅ Already know it
Flow + automation thinking ✅ Already know it
Debugging + org setup ✅ Already know it
Topic + Instruction writing ⬅ This is the only new muscle