Section 4 · AI Agents · 35% of exam
4.8 Configure Topics, Instructions, and Actions for a Use Case
Exam objective
Given a business requirement, configure focused topics or subagents, instructions, and actions so the agent can complete the intended task.
Translate a request into a design
Start by defining the user, goal, data needed, boundaries, and acceptable outcome. Break a broad job into focused responsibilities that can be routed clearly. For each topic or subagent, decide what requests belong there, what information is needed, and which actions can fulfill the task. Do not start by adding every available action. Map each capability to a real requirement and identify where the agent must ask for missing details, verify identity, or escalate. A simple design is easier to test and safer to maintain.
Write layered instructions
The agent’s top-level instructions set global behavior such as its role, tone, or broad guardrails. Topic or subagent instructions define how to handle a particular job. Action instructions describe when and how to use an individual operation, including its inputs and outputs. Keep these layers consistent. Avoid copying a long generic paragraph into every topic. State essential boundaries plainly, such as which source to use, which facts to confirm, what not to promise, and what to do when an action returns no result. Instructions guide reasoning; filters and deterministic logic enforce conditions that must always hold.
Connect actions to the workflow
For a customer asking about an order, the agent may need a lookup action, a cancellation action, and an escalation path. The lookup action should receive a reliable order identifier and return status. The cancellation action should not be exposed until policy and authorization conditions are satisfied. Map important results to variables, use filters for conditional availability, and specify transitions or follow-up actions when order matters. Ensure that each action is assigned to the appropriate topic or subagent. Keep action descriptions distinct so the model can select the right operation rather than choosing between near-duplicates.
Validate the configuration
Test the design with realistic utterances, paraphrases, ambiguous requests, and out-of-scope messages. Observe which topic is selected, whether the agent requests missing data, whether filters allow the expected action, and whether action outputs are interpreted correctly. Test prohibited paths too, such as an unverified user requesting a sensitive update. Check that the response reflects the actual action result and does not invent completion. If routing is wrong, refine classification descriptions and scope. If the wrong action runs, review its description, the assigned actions, variables, and any deterministic conditions. Iterate on the responsible layer instead of rewriting all instructions at once.
TIP: Design backward from the desired outcome: list required facts and actions, then configure routing and controls that ensure the agent follows the appropriate path.
A topic’s name or description alone does not authorize an action. Access controls and the underlying automation’s permissions must also allow the operation.
Extra concept: The agent’s Preview trace helps connect observed behavior to specific routing and action decisions during configuration review.
Official references:
- Configure Your Agent https://help.salesforce.com/s/articleView?id=ai.agent_parent_configure.htm&language=en_US&type=5
- Subagent Classification and Routing https://help.salesforce.com/s/articleView?id=ai.agent_topics_routing.htm&language=en_US&type=5
- Explore Agentforce Builder https://help.salesforce.com/s/articleView?id=ai.agent_builder_tour.htm&language=en_US&type=5
Key takeaways
- Translate requirements into a focused routing and action design.
- Keep agent, topic, and action instructions consistent and specific to their responsibilities.
- Test correct, ambiguous, missing-data, denied-access, and failure scenarios, then adjust the responsible component.
