Overview
Testing, Deployment, and Maintenance is worth 10% of the exam and has four objectives. Two cover Agentforce Testing Center: running tests against an agent and understanding how its evaluations score results. Two cover application lifecycle management, deploying an agent and deploying a prompt template from a sandbox to production.
In Testing Center you create a test with test cases that pair an utterance with the expected topic, the expected actions and the expected response. Test cases can be written by hand, uploaded from a CSV file or generated with AI from the agent's topics. Evaluations compare actual behavior with those expectations, using exact checks for topics and actions and an AI-based judgment for response quality. Deployment questions focus on metadata and dependencies: the agent, its topics and actions, the flows, Apex and prompt templates behind them, the agent user, channel setup and any Data 360 components that must exist in the target org.
How it's tested
Questions describe a release in progress: a test run with failures, a team preparing to move an agent or template to production, or a deployment that broke. You interpret the evaluation, choose the right test approach, or name the dependency or post-deployment step that was missed.
-
Read the failed test
A test case fails on topic, action or response, and you identify which part of the agent to change.
-
Build the test set
A team needs broad coverage quickly, and you choose AI-generated cases, CSV upload or hand-written cases for edge conditions.
-
Deployment dependencies
An agent or template works in sandbox but fails in production, and you identify the missing component, activation or permission.
After deploying a Service agent to production through a change set, the order-status action fails for every customer. In the sandbox it worked. The flow behind the action deployed, but the permission set granting the agent user access to Order fields was not included, and the agent user in production is a different user. Deploying and assigning the permission set to the production agent user resolves it.
Objectives 4
Official objectives from the Salesforce outline, in outline order.
4.1 Given a scenario, test an agent using Testing Center.Creating and running agent test sets.
- What it means
- Agentforce Testing Center runs batches of test cases against an agent. Each case contains an utterance and the expected topic, actions and response. Cases can be entered manually, uploaded in a CSV template or generated by AI from the agent's configuration, and runs are best done in a sandbox.
- What to understand
- Use AI-generated cases to get broad coverage quickly, then add hand-written cases for edge conditions, policy limits and adversarial phrasing. Keep expected values precise so failures are meaningful. Tests can execute real actions and consume usage, which is another reason to run them in a sandbox with test data. Re-run the same test after each change to catch regressions. Tests can also be run from the command line with Agentforce DX for automation in pipelines.
- How it shows up
- A team must validate an agent before a tight release date, and you choose how to build and run the test set.
- Common confusion
- Running tests in production where actions can change real customer records.
4.2 Explain how Testing Center evaluations work.Scoring topics, actions and responses.
- What it means
- Testing Center evaluates each test case against its expectations. Topic and action checks compare what the agent selected with what was expected. Response evaluation uses an AI judge to score whether the actual answer matches the expected answer's meaning, and additional evaluations can check factors such as completeness or latency.
- What to understand
- Read results by layer. A topic mismatch suggests classification descriptions or overlapping scope. An action mismatch suggests action descriptions, filters or missing variables. A low response score with correct topic and action suggests instructions, grounding or the action output itself. Because the response check is semantic, wording need not match exactly; expected responses should describe the essential content rather than one exact phrasing.
- How it shows up
- A test result shows which checks passed and failed, and you decide what to change in the agent.
- Common confusion
- Expecting the response evaluation to require word-for-word matches with the expected answer.
4.3 Identify the considerations for deploying an agent from sandbox to production.Agent metadata, dependencies and post-deployment setup.
- What it means
- An agent is represented in metadata as a bot and bot version with planner, topic and action components, and it depends on flows, Apex classes, prompt templates, permission sets and channel configuration. Deployment can use change sets, the Metadata API, Salesforce CLI or DevOps Center.
- What to understand
- Include every dependency or confirm it already exists in production. The agent user is a user record, not metadata, so it must be created and assigned the right permission sets in the target org. Data 360 assets such as search indexes and retrievers move through Data 360 data kits or are rebuilt, not through standard change sets alone. After deployment, confirm channel settings and Omni-Channel routing, activate the agent and run a smoke test.
- How it shows up
- An agent behaves differently after deployment, and you identify the missing dependency or setup step.
- Common confusion
- Assuming the agent user and its permission assignments deploy automatically with the agent.
4.4 Identify the considerations for deploying a template from sandbox to production.Moving prompt templates with versions and grounding.
- What it means
- Prompt templates are metadata that can be deployed with change sets, the Metadata API or the CLI. A template carries its versions, but the flows, Apex classes, fields and retrievers it references must also exist in the target org.
- What to understand
- Deploy grounding dependencies with or before the template. Check which version is active after deployment and activate the intended one in production. Model configurations and Data 360 retrievers referenced by the template must exist in the target org with matching names. Permission set assignments for Prompt Template User and Manager are org-specific. Preview the template against production records before users rely on it.
- How it shows up
- A template deploys but fails or uses the wrong version in production, and you name the cause.
- Common confusion
- Deploying only the template while leaving its prompt flow or Apex grounding class behind.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Agentforce Testing Center
Tool for batch testing agents with expected topics, actions and responses.
- Remember
- Run it in a sandbox with test data.
- Tested as
- Every testing question.
- Trap
- Testing only by chatting manually.
-
Test case
An utterance with expected topic, actions and response.
- Remember
- Precise expectations make failures meaningful.
- Tested as
- Building test sets.
- Trap
- Leaving expected actions blank.
-
AI-generated test cases
Cases created automatically from the agent's topics and actions.
- Remember
- Fast broad coverage, then add edge cases.
- Tested as
- Time-pressured release scenarios.
- Trap
- Relying on them alone for edge conditions.
-
CSV upload
Importing test cases from a template file.
- Remember
- Reuses cases written by business experts.
- Tested as
- Large existing test libraries.
- Trap
- Ignoring required columns.
-
Response evaluation
AI judge scoring the meaning of the actual response.
- Remember
- Semantic, not word-for-word.
- Tested as
- Interpreting response failures.
- Trap
- Writing expected answers as exact phrases.
-
Regression testing
Re-running the same test after every change.
- Remember
- Catches new failures caused by fixes.
- Tested as
- Ongoing maintenance.
- Trap
- Testing only the changed topic.
-
Agent metadata
Bot, bot version, planner, topic and action components.
- Remember
- Deploy with flows, Apex, templates and permission sets.
- Tested as
- Deployment dependency questions.
- Trap
- Expecting one component to carry everything.
-
Agent user setup
Creating and assigning permissions to the production agent user.
- Remember
- It is data, not metadata.
- Tested as
- Post-deployment failures.
- Trap
- Expecting it to deploy.
-
Data 360 data kits
Packages for moving Data 360 configuration between orgs.
- Remember
- Used for indexes, retrievers and related assets.
- Tested as
- Agents that rely on retrieval.
- Trap
- Using a change set alone.
-
Template version activation
Making the intended version active in the target org.
- Remember
- Check after deployment.
- Tested as
- Old output after release.
- Trap
- Assuming the latest version auto-activates.
-
Agentforce DX
Developer tooling and CLI commands for agents and tests.
- Remember
- Supports automated testing in pipelines.
- Tested as
- Source-driven development scenarios.
- Trap
- Assuming tests can run only in the UI.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Reading test failures
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Topic mismatch | Wrong topic selected | Fix classification descriptions or scope | Editing action code first |
| Action mismatch | Wrong or missing action | Fix action descriptions, filters or variables | Rewriting the topic |
| Low response score | Answer meaning differs | Fix instructions, grounding or action output | Changing the expected answer to match |
Ways to build test cases
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| AI-generated | Created from agent configuration | Fast breadth | Missing edge cases |
| CSV upload | Imported from a file | Reuse expert-written cases | Wrong column format |
| Manual entry | Written in Testing Center | Precise edge and negative cases | Too few cases for coverage |
Deployment items
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Agent and topics | Bot and planner metadata | Deployable metadata | Leaving actions behind |
| Actions' flows and Apex | Logic behind actions | Must deploy first or together | Assuming they already exist |
| Agent user | User record | Created and assigned in target | Expecting it in a change set |
| Data 360 retrieval assets | Indexes and retrievers | Data kits or rebuild | Using change sets only |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Test in production
- What candidates pick
- Run the full test suite in production to get realistic results.
- Why it looks right
- Production has real data and configuration.
- Why it's wrong
- Tests execute actions and consume resources; a sandbox with test data avoids changing customer records.
Clue to watch for Actions update or create records.
-
Edit the expectation
- What candidates pick
- Change the expected response so the test passes.
- Why it looks right
- The test turns green quickly.
- Why it's wrong
- It hides a real defect; investigate whether instructions, grounding or action output caused the difference.
Clue to watch for The expected answer reflects the business requirement.
-
One-click agent deploy
- What candidates pick
- Deploy only the agent metadata.
- Why it looks right
- The agent appears as one item in Setup.
- Why it's wrong
- Actions depend on flows, Apex and templates, and the agent user and Data 360 assets need separate handling.
Clue to watch for Actions fail after deployment.
-
Exact phrase matching
- What candidates pick
- Write expected responses as exact sentences.
- Why it looks right
- Exact checks seem rigorous.
- Why it's wrong
- Response evaluation judges meaning, so expectations should describe required content rather than one phrasing.
Clue to watch for Responses vary in wording but are correct.
-
Latest version is live
- What candidates pick
- Assume the newest deployed template version runs.
- Why it looks right
- New usually means active.
- Why it's wrong
- Only the active version runs; deployment does not guarantee the intended version is activated in the target org.
Clue to watch for Users see old output after release.
Answer framework
Use these steps when a scenario covers testing or releasing agents and templates.
- Build a test set in a sandbox, combining AI-generated cases for breadth with hand-written cases for edge and negative conditions.
- Run the tests and read failures by layer: topic, then action, then response, fixing the earliest failing layer first.
- Re-run the full test after each fix to catch regressions before planning the deployment, since an instruction change for one topic can shift routing for another.
- List every dependency, including flows, Apex, prompt templates, permission sets, channel settings and Data 360 assets, and choose how each will move.
- After deployment, set up the agent user, activate the agent and intended template versions, and run a smoke test in production through a real channel conversation.
Practice question
Want more like this? The AI-201 practice test has questions for every domain.
Checklist
0 of 11
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Spend most of your time in Testing Center rather than reading about it. Take an agent you built for the AI Agents section and create a test with AI-generated cases, then add five hand-written cases that target edge conditions, such as an unverified customer or an out-of-scope request. Run it, change one topic description and one action description, and run it again to see how results shift. Download the CSV template and upload a small set of cases so you know its columns.
For deployment, create a change set or a Salesforce CLI project that includes an agent, its flow and Apex actions, a prompt template and a permission set. Note which items the dependency list adds automatically and which it misses, then write your own checklist that includes the agent user and Data 360 assets.
High-priority ideas are reading failures by layer and agent dependencies. Underestimated areas are template version activation and the semantic nature of response scoring. Use Salesforce Help pages on Testing Center and agent deployment, since this area changed across 2025 and 2026 releases.
FAQ
What share of the exam is Testing, Deployment, and Maintenance?
Testing, Deployment, and Maintenance is 10% of the exam per third-party copies of the current guide, around six scored questions as a planning estimate, across four objectives.
Did this section shrink from the previous outline?
Yes. The 2025 outline had a 20% Development Lifecycle section that included adoption monitoring. Monitoring now sits under Governance and Observability, and this section focuses on testing and deployment.
Is Testing Center the same as Agentforce Testing Center?
Yes. The guide uses the shorter name. Salesforce documentation refers to Agentforce Testing Center, and Agentforce DX provides command-line access to tests.
Can I deploy agents with change sets?
Yes, along with the Metadata API, Salesforce CLI and DevOps Center. Include dependencies, and set up the agent user and Data 360 assets separately.
Why does an agent need a smoke test after deployment?
Production can differ in users, permissions, data and channel routing. A short test confirms that the deployed agent routes, acts and responds correctly before customers use it.
Does deploying a prompt template also deploy its grounding flow?
Not automatically in every tool. Add the flow, Apex classes and referenced fields to the deployment or confirm they already exist in production.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 4 knowledge
You have covered Testing Center, evaluations and moving agents and templates to production. Check that you can read a failing test and spot a missing dependency quickly.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AI-201 AI tutor Get any concept explained your way