Agentforce Specialist testing and governance
Testing, Deployment, and Maintenance (10%) and Governance and Observability (10%) are 20% of the Salesforce Agentforce Specialist exam between them — around 12 of the 60 scored questions.
Two modest sections covering what happens either side of release: proving the agent works before it ships, and knowing whether it still works afterwards.
Testing, deployment and maintenance (10%)
Testing Center
Salesforce’s tool for testing agents, and the outline asks both how to use it and how its evaluations work.
The idea to internalise: an agent has no single right answer, so you cannot unit-test it the way you would test code. Instead you assemble representative scenarios with expected outcomes and evaluate the agent against them repeatedly.
That matters most at change time. When a question asks how to confirm a modified agent has not got worse, the answer involves Testing Center evaluations compared across versions — never asking colleagues to try it, and never releasing and watching for complaints.
Deploying an agent from sandbox to production
Standard Salesforce release management applied to agents, with agent-specific considerations: what moves with the agent, what must exist in the target org first, and what has to be reconfigured after deployment.
The recurring trap links to the AI Agents section: the security context differs between orgs. An agent that executed actions successfully in a sandbox may fail in production because the running identity’s permissions differ.
Deploying a prompt template from sandbox to production
Listed as a separate objective from deploying an agent, which is a deliberate signal that the considerations differ. Know both, and remember that a deployed prompt template still needs activating in the target org.
Governance and observability (10%)
Managing and monitoring agents
Knowing what agents exist, who owns them, and what they are doing in production. An inventory problem as much as a technical one — ungoverned agents proliferate the same way ungoverned reports always have.
Agent analytics and optimisation
What analytics reveal and what you do with them:
- What users actually ask, which is rarely what you designed for
- Where the agent fails to resolve a request or escalates
- Which topics and actions are used, and which never fire
- Patterns showing where a new topic or a better grounding source is needed
The loop the exam expects: observe what happens, find where it falls short, improve topics, actions or grounding, and re-test. Optimisation is informed by analytics rather than by intuition.
How the two connect
A well-run agent is tested before release and observed after it. Questions sometimes span both, describing a team that ships an agent and has no idea whether it is helping anyone.
Sample questions
Question 1. A team has modified an agent's topics and grounding and must demonstrate to stakeholders that response quality has not regressed. What should they do?
- A. Ask several colleagues to try the agent and report their impressions
- B. Run the existing Testing Center evaluations against both versions and compare
- C. Release it and watch agent analytics for a fortnight
- D. Add more content to the Data Library before release
Show answer
Answer: B
Running the same Testing Center evaluations against both versions produces a comparison that can be shown, which is what demonstrating no regression requires. Informal colleague testing is not repeatable, releasing and monitoring finds regressions after users experience them, and additional grounding does not evidence anything.
Question 2. A prompt template was deployed from sandbox to production, but users in production cannot use it. The deployment itself reported success. What is the most likely cause?
- A. The Data Library was not deployed alongside it
- B. The agent’s channel configuration is incorrect
- C. Model access management is blocking the model
- D. The template has not been activated in production
Show answer
Answer: D
A prompt template must be activated in the target org before it can be executed, and a successful deployment does not activate it. A Data Library relates to grounding, channel configuration concerns agents, and model access would produce a different failure.
Question 3. Agent analytics show that a quarter of conversations end with the agent escalating on the same class of question. What is the appropriate response?
- A. Add or improve a topic and its grounding to cover that class of question
- B. Disable escalation so conversations do not end that way
- C. Restrict the agent to fewer users
- D. Switch the agent to a different model
Show answer
Answer: A
Analytics revealing a consistent gap point to a missing capability, so adding a topic or improving grounding for that question class addresses the cause. Disabling escalation hides the failure from users, restricting access reduces value, and a different model does not supply missing capability or content.
What to practise
Build a small set of Testing Center evaluations for an agent, change something deliberately, and re-run them to see the difference reported.
Then deploy a prompt template to another org and try to use it before activating it. Both of these take minutes and both cover objectives that are stated separately in the outline precisely because people get them wrong.