DESIGNER GUIDE

Test and troubleshoot your solution

Test workers directly, run complete customer journeys, investigate incomplete configuration and confirm that a project is ready before publishing or deployment.

Solution designer15–20 minutes

Scope: draft worker or Flow project

Test behaviour, not only appearance

A solution is not ready because the first card renders correctly. Confirm the worker’s decisions, every important flow branch, missing information, failures and the completed end-to-end experience.

Three levels of testing

Each level confirms something the others cannot.

Levels of testing
LevelWhat it confirms
Direct worker testThe worker interprets the request and returns the expected response contract
Integrated flow testThe flow passes data to the worker or step and handles the result correctly
End-to-end Run DemoThe complete user journey behaves correctly from the first interaction to completion
One level is never enoughA successful direct worker result does not prove that the containing flow maps or branches on that result correctly.

On this page

Before you begin

Prepare the draft, the data and the expected results before running the first test.

  • Save the current draft.
  • Use fictional test data.
  • Define the expected outcome before each test.
  • Prepare happy-path, incomplete and invalid examples.
  • Identify every important branch.
  • Confirm the expected worker-response structure.
  • Reset conversational state between independent scenarios.
  • Record the project version being tested.
  • Avoid testing production actions from a draft environment.
Write expected results firstA test is meaningful only when the designer knows what should happen. Record the expected intent, branch, message, action and completion state before running it.

1. Prepare a test matrix

A short matrix keeps coverage honest and repeatable across design changes.

Example test matrix
ScenarioExampleExpected result
Clear supported request“Where is my package?”Late-delivery or order-tracking intent is selected
Alternative wording“My parcel should have arrived yesterday”The same approved process is selected
Missing information“My delivery is late” without an order numberThe solution asks for the required information
Ambiguous request“There is a problem with my order”The worker requests clarification or returns uncertainty
Unsupported requestA question outside customer supportApproved fallback, main menu or human support
Invalid inputAn incorrectly formatted order referenceControlled validation message
Sensitive actionA request requiring approvalHuman approval or approved escalation path
Adversarial inputAn attempt to override the worker’s roleInstalled guardrails and instructions prevent unsafe behaviour
Test the wording customers actually useInclude spelling mistakes, abbreviations, short requests and messages containing more than one issue.

2. Test an Agentic Worker directly

Start with the worker in isolation so interpretation problems are not confused with flow problems.

The Test panel provides a clean conversation and suggested starting messages.
Inspect whether the worker returns the intended classification and response structure.
Reset chat
Clears the current test conversation so an independent scenario does not inherit earlier context.
Suggested messages
Quick starting examples. They are not a complete test set.
Message input
Accepts the test request.
  • The user enters an open customer question.
  • The worker identifies a likely late-delivery intent.
  • The response includes structured routing information.
  • The result names the process intended to handle the request.
  • The exact fields shown are specific to this demonstration.
  1. 1Open the worker in Composer.
  2. 2Select the Test panel.
  3. 3Select Reset chat.
  4. 4Enter one test message.
  5. 5Compare the returned result with the expected contract.
  6. 6Confirm the detected intent.
  7. 7Confirm extracted values and uncertainty handling.
  8. 8Confirm that the worker does not attempt the deterministic resolution itself.
  9. 9Repeat the test with alternative wording.
  10. 10Test a missing-information request.
  11. 11Test an unsupported request.
  12. 12Reset the chat before each independent scenario.
Test the response contractA result may be understandable to a human but unusable by a flow when a required field is missing, renamed or surrounded by extra text.

3. Test the complete flow in Run Demo

Run Demo tests the current flow through the same structured interactions the user will experience.

Run Demo tests the current flow through the same structured interactions the user will experience.
  • Initial content and instructions
  • Buttons and structured choices
  • Open-text input
  • Required fields
  • Validation messages
  • Worker routing
  • Called flows
  • Completion messages
  • Fallback behaviour
  1. 1Save the current draft.
  2. 2Open Run Demo.
  3. 3Begin with the main happy path.
  4. 4Complete each displayed input.
  5. 5Confirm the expected next step appears.
  6. 6Complete the entire journey.
  7. 7Restart the interaction.
  8. 8Test an alternative branch.
  9. 9Test missing or invalid data.
  10. 10Test an open-text request that invokes the worker.
  11. 11Confirm that the correct deterministic flow starts.
  12. 12Repeat until every important route has been covered.
Complete the journeyDo not stop after confirming that a button works. Continue until the process reaches its expected completion or approved escalation.

4. Test worker-to-flow routing

Interpretation, integration and the end-to-end process each need their own confirmation.

1. Test the worker directly
       ↓
2. Test the Agentic Worker step in the flow
       ↓
3. Test the complete process from Run Demo
Stage 1: Worker
Confirm that the request produces the expected intent and structured fields.
Stage 2: Flow integration
Confirm that the Agentic Worker step passes the correct input and that the following branch receives the expected result.
Stage 3: End-to-end process
Confirm that the customer request starts the intended deterministic resolution flow and completes successfully.
  1. 1Enter “Where is my package?” in the worker Test panel.
  2. 2Confirm the worker returns a late-delivery classification.
  3. 3Enter equivalent wording in the customer-facing Run Demo.
  4. 4Confirm the main flow invokes the router.
  5. 5Confirm the late-delivery branch starts.
  6. 6Confirm the deterministic resolution requests the required information.
Do not assume direct success guarantees integration successThe worker may classify correctly while the flow checks the wrong field, expects a different value or sends incomplete input.

5. Investigate a project issue

An issue count is a signal to review required configuration and connections.

The project toolbar can show an issue count while the canvas is being configured.
Select relevant steps and review their required settings and references.
Read the indicator carefullyThe toolbar shows that the project currently has one issue. The interface shown does not display the issue’s detail or identify the affected node, so do not assume that the selected step caused it. Use the count as a prompt to review required configuration and connections.
  1. 1Save the current project.
  2. 2Note the displayed issue count.
  3. 3Inspect recently added or changed steps.
  4. 4Confirm that every required field has a value.
  5. 5Confirm that referenced workers, flows and tools are available.
  6. 6Check that process steps have the required incoming and outgoing connections.
  7. 7Check branch outcomes and destinations.
  8. 8Confirm that expected variables are present.
  9. 9Run Demo again.
  10. 10Continue to the readiness check after the issue is resolved.
Common configuration checks
CheckExample
Required resourceAgentic Worker step has an available worker selected
Required destinationGo to Flow points to an available flow
Required capabilityMCP, HTTP or extension resource is available
ConnectionThe step is linked to the intended next route
BranchEvery expected result has a controlled destination
InputRequired data is passed into the step
OutputThe following step expects the fields actually returned
Do not fix issues by deleting validationCorrect the missing configuration or process design rather than bypassing the control that detected it.

6. Confirm readiness before shipping

The readiness check confirms when Designer found no blocking configuration issues.

The readiness check confirms when no blocking configuration issues were found.
  • A successful result states that Designer found no issues.
  • The summary may show cards and step counts.
  • Let’s go continues toward publishing or deployment.
  • A successful technical result should be combined with business testing.
  1. 1Resolve the current issue count.
  2. 2Save the draft.
  3. 3Run each priority test.
  4. 4Record the test result.
  5. 5Continue toward deployment.
  6. 6Review the readiness result.
  7. 7Return to the Editor if a blocking issue remains.
  8. 8Continue only after both technical validation and business review succeed.
Ready to ship is not the same as approved to go liveA readiness check does not replace business-owner review, customer acceptance, security review, approval-policy review, integration testing or production change controls.

The next stage is covered in the Publishing and deployment guide.

7. Retest after a change

A correction is only complete when the failed scenario passes and nearby scenarios still pass.

Reproduce
   ↓
Change one thing
   ↓
Save
   ↓
Run the failed scenario
   ↓
Run a nearby successful scenario
   ↓
Run validation
   ↓
Save as version
  1. 1Reproduce the problem before changing the design.
  2. 2Make one focused correction.
  3. 3Save the draft.
  4. 4Repeat the failed test.
  5. 5Repeat at least one related passing test.
  6. 6Check that the correction did not break another branch.
  7. 7Run the readiness check.
  8. 8Save a meaningful version when the correction is accepted.
Retest nearby pathsA change to one branch, worker contract or variable can affect other paths that consume the same data.

Testing checklist

Confirm this list before a solution moves toward publishing or customer acceptance.

  • Expected results were defined before testing.
  • Only fictional test data was used.
  • The worker was tested directly.
  • Independent worker tests used Reset chat.
  • The happy path completed successfully.
  • Alternative wording was tested.
  • Missing information was tested.
  • Invalid input was tested.
  • Ambiguous requests were tested.
  • Unsupported requests followed the approved fallback.
  • Important branch boundaries were tested.
  • Worker output matched the flow contract.
  • Called flows started correctly.
  • Human approval paths were considered.
  • The visible issue count was resolved.
  • A readiness check completed successfully.
  • Business acceptance was recorded separately.
  • The accepted project state was saved as a version.

Testing troubleshooting

Frequent testing problems and the practical action that resolves them.

Testing troubleshooting
ProblemLikely causeAction
A direct worker test is influenced by an earlier messageThe test conversation still contains previous contextSelect Reset chat before the independent scenario
The worker chooses the wrong intentInstructions or intent boundaries are unclearRefine the definitions and test alternative wording
The worker returns extra prose around structured outputThe response contract is too looseRequire only the specified structure and retest
The worker is correct but the wrong flow branch startsThe flow checks the wrong field or valueCompare the worker output with the branch condition
Run Demo waits without continuingThe current step requires user input or a button actionComplete the displayed interaction and verify the intended next step
A step has no available resourceThe required worker, flow or integration is unavailableConfigure, publish or expose the referenced resource
The toolbar shows an issueOne or more required settings or connections are incompleteReview recently changed steps and required fields
The readiness check failsA blocking project issue remainsReturn to the Editor and correct the reported configuration
The readiness check passes but users receive the wrong outcomeTechnical validation did not cover the business ruleExpand the test matrix and correct the process design
A correction breaks another branchShared variables or connections changedRetest nearby paths and restore the intended mappings
Test results vary too muchWorker instructions or model settings allow excessive variabilityTighten the contract and review model configuration
Real systems were affected during a draft testProduction resources were connected to the test processStop testing, review integration scope and use a safe test environment
Back to Designer Guides