AI Agents

Splitting into two agents with Platform Event coordination is the correct architecture pattern. The first agent handles synchronous Salesforce operations (Flow execution, record updates) ensuring immediate data consistency. The second agent handles the external API call asynchronously, with Platform Events providing durable messaging between agents. This pattern isolates timeout failures to the external integration without blocking the primary workflow, and Platform Events provide guaranteed delivery and retry capabilities.

Adding explicit positive examples in the correct Topic and negative examples in the competing Topic directly addresses the classification confusion. Agentforce uses the Topic instructions, including examples, as context for the LLM during intent routing. Positive examples ('wireless keyboards' specifically) strengthen the 'Keyboards & Mice' Topic signal, while negative examples in 'Computer Accessories' explicitly exclude the overlapping category, creating clearer boundaries between Topics. This is the standard pattern for resolving Topic routing ambiguity.

System Context with custom Apex filtering is the correct approach for this security requirement. The agent needs to access restricted schedule data (bypassing field-level security) but must still enforce business logic about which records patients can see. By running in System Context, the agent can read all Provider_Schedule__c records including restricted fields. The custom Apex action then implements the insurance-eligibility business rule to filter records appropriately before returning results. This separates technical security (FLS) from business logic (insurance matching), which is the recommended pattern for complex security scenarios.

Pagination is the correct architectural pattern for handling large result sets in agent actions. By adding 'limitResults' and 'offsetResults' input parameters, the agent can request contacts in manageable chunks of 50 records per call. The action returns metadata (total count, current page) allowing the agent to inform the user ('Showing 50 of 150 contacts') and offer to load more. This respects the response size limits while still providing complete data access. Pagination is a standard pattern for Agentforce actions dealing with large datasets.

Adding a date-based instruction in the Topic configuration directly tells the LLM to filter grounding results by article recency during retrieval. When the agent performs grounding searches via Einstein Trust Layer, the instruction becomes part of the retrieval prompt, causing the system to prioritize or exclusively use recent articles. This is the simplest and most direct method that leverages the agent's natural language understanding to apply business rules (recency requirements) to grounding queries. The '30 days' window provides buffer beyond the weekly update cycle.

Using native agent action chaining with variable mapping is the correct declarative approach. The external API action's output can be configured to extract specific JSON fields (trackingNumber) into a variable within the Topic context. A subsequent Update Record action then references that variable to populate Case.Return_Tracking__c. This approach uses Agentforce's built-in capabilities without custom code, maintains the agent's understanding of the workflow (for conversation continuity), and keeps the operation synchronous so the agent confirms the tracking number was stored before responding to the customer.