Section 5 · Testing, Deployment, and Maintenance · 10% of exam
5.1 Testing Agents with Testing Center
Exam objective
Given a scenario, plan and run an agent test in Testing Center using representative cases and appropriate expected behavior.
Why test beyond one conversation
Agentforce responses can vary because the reasoning engine interprets language and context. A successful chat in Preview does not show whether the agent will respond reliably across other phrasings, missing details, or failure conditions. Testing Center supports batch testing so a builder can evaluate many scenarios together before releasing an agent. It helps check routing, action selection, and responses at a larger scale than manually entering one utterance at a time. Use Builder Preview for quick investigation and Testing Center for repeatable suites that can reveal patterns and regressions.
Create a useful test suite
A test suite defines the agent version, cases, and evaluation approach. Cases can be written manually, generated from the agent’s subagents and actions, created from knowledge sources, or uploaded from a file; current Testing Center versions also support API-driven workflows. A meaningful case describes the user utterance and the expected subagent, action, and response or outcome. Include ordinary requests, alternative phrasings, multi-turn interactions, clarification requests, unsupported or out-of-scope asks, and situations where an action fails. For a customer-facing agent, include authentication and permission boundaries. Match coverage to the agent’s actual deployment, user types, and risk.
Run and inspect tests
Run the suite against the intended agent version and confirm that all cases completed. Review the per-case results rather than relying on an overall score. A failed subagent assertion can indicate a routing description or scope issue. An action assertion can reveal an action-selection, instruction, filter, or prerequisite issue. A weak response can point to missing context, outdated knowledge, or unclear instructions. Use the detail view and interaction trace to see what happened. Retest a failed utterance in Builder to reproduce the behavior, make a focused correction, and run the relevant suite again.
Use tests as a repeatable asset
Save important cases so they can be rerun after changes to instructions, actions, data, channels, or models. Add a new case when a production issue reveals an uncovered scenario. Keep expected results realistic: a language model may word a correct response differently, so define the facts and outcome that matter rather than requiring identical phrasing. A test suite provides evidence for a release decision, but it cannot cover every possible conversation. Prioritize high-impact tasks, common user needs, access boundaries, and failure modes. Record the tested agent version and the result so later changes can be compared against a known baseline.
TIP: Test more than the happy path. Include paraphrases, missing information, multi-turn context, out-of-scope questions, action failures, and access restrictions.
Testing Center has a newer Agentforce Studio experience and a legacy Setup experience. Available suites, scorers, and labels can differ, so verify the interface and feature availability in your org. Extra concept: A reusable suite turns a one-time investigation into a regression check that can be run again after each meaningful change.
Official references:
- Test Agents in Testing Center https://help.salesforce.com/s/articleView?id=ai.agent_studio_testing_center_test_agents.htm&language=en_US&type=5
- Set Up and Run Tests in Agentforce Studio https://help.salesforce.com/s/articleView?id=ai.agent_studio_testing_center_setup_tests.htm&language=en_US&type=5
- Build an Agent Testing and Observability Plan https://help.salesforce.com/s/articleView?id=ai.agent_test_prep.htm&language=en_US&type=5
Key takeaways
- Testing Center runs many agent scenarios as a batch and supports reusable test suites.
- Define expected routing, actions, and outcomes for representative utterances and conversations.
- Use detailed failures to find the responsible configuration layer, then retest after the fix.
