Monitor

Agent Testing

Build confidence in your agent's behavior with automated testing

Agent testing lets you verify conversational responses, tool usage, and full multi-turn outcomes before you deploy. Create tests from scratch or from existing conversations, then run them from the dashboard, CLI, or API.

Video Walkthrough

Overview

The framework includes three complementary test types:

  • Simulation Testing — Runs end-to-end, multi-turn conversations with a simulated user
  • Next Reply (Scenario) Testing — Validates the agent’s next response against success criteria
  • Tool Call Testing — Ensures the agent calls the right tool with the right parameters

When to use which test

Test typeUse when you need to
SimulationCheck that a full conversation reaches a defined outcome
Next Reply (Scenario)Check that the agent’s next message meets quality, tone, or policy criteria
Tool CallCheck that the agent invokes a specific tool with expected parameters

Creating tests from conversations

Transform real conversations into test cases when you find an interaction where the agent underperformed.

Creating test from conversation
  1. Open the conversation in call history
  2. Click Create test from this conversation
  3. Review the prefilled context, then define the expected behavior
  4. Add the test to your suite to catch similar failures later

Simulation Testing

Simulation testing evaluates your agent across a full, multi-turn conversation with a simulated AI user. Unlike Next Reply tests, this type checks whether the complete interaction reaches your defined outcome.

Creating a Simulation Test

Simulation test creation UI
1

Define the scenario

Describe the user’s context, intent, and behavior in natural language. The simulator uses this scenario to drive the conversation.

Example scenario:

“A tourist who is not fluent in English is trying to place an order at a restaurant.”

2

Set the success condition

Define the outcome that should count as a pass. This prompt is used to evaluate whether the full conversation succeeded.

Example success condition:

“The agent confirmed the order details, handled clarifying questions, and completed the order without misunderstandings.”

3

Set max turns

Choose how long the simulation can run before stopping. Use a lower value for focused checks and a higher value for complex workflows.

  • Minimum: 1
  • Maximum: 50
  • Default: 5
4

Run and review the result

Execute the test and inspect the generated conversation transcript. Review the pass/fail result against your success condition, then iterate on your prompt, tools, or agent configuration.

Optional Configuration

You can refine simulation behavior in the test configuration panel:

  • Environment: Select which environment to test against when your agent has multiple environments configured. If only one environment is available, this selector is hidden.
  • Chat history: Start from a partial conversation instead of a blank state. This is useful for testing in-progress conversations and recovery behavior.
  • Dynamic variables: Inject test-specific values into your agent variables (for example, user names or order IDs) without changing the base agent configuration.

Tool Mocking

Simulation tests support tool mocking so your agent can receive controlled responses during a run instead of calling live systems.

Mocking strategy

  • Mock none: No tools are mocked.
  • Mock all tools: Every mockable tool returns a mock response.
  • Mock selected tools: Only tools you explicitly choose are mocked.

System tools and workflow tools are never mocked.

Fallback behavior

If a mocked tool is called and no matching mock response is found, choose one of these behaviors:

  • Call real tool: Executes the real tool call.
  • Finish with error: Returns an error response from the tool instead of calling the real tool.

The fallback setting appears only when at least one tool is mocked.

Next Reply (Scenario) Testing

Next Reply (Scenario) testing evaluates only the agent’s next message, not a full multi-turn outcome. Provide conversation history that leads up to the reply you want to evaluate, then score that reply against success criteria.

For full multi-turn outcomes, use Simulation Testing.

Creating a Next Reply Test

Next Reply (Scenario) Testing Interface
1

Define the chat history

Provide the conversation history leading up to the reply you want to evaluate. This can be a single user message or multiple turns of context.

Example chat history:

User: "I'd like to cancel my subscription. I've been charged twice this month and I'm frustrated."
2

Set success criteria

Describe in plain language what the agent’s response should achieve. Be specific about the expected behavior, tone, and actions.

Example success criteria:

  • The agent should acknowledge the customer’s frustration with empathy
  • The agent should offer to investigate the duplicate charge
  • The agent should provide clear next steps for cancellation or resolution
  • The agent should maintain a professional and helpful tone
3

Provide examples

Supply both success and failure examples to help the evaluator understand the nuances of your criteria.

Success example:

“I understand how frustrating duplicate charges can be. Let me look into this right away for you. I can see there were indeed two charges this month - I’ll process a refund for the duplicate charge immediately. Would you still like to proceed with cancellation, or would you prefer to continue once this is resolved?”

Failure example:

“You need to contact billing department for refund issues. Your subscription will be cancelled.”

4

Run the test

Execute the test. An LLM evaluator compares the agent’s next reply against your success criteria and examples to determine pass/fail status.

Tool Call Testing

Tool call testing verifies that your agent correctly uses tools and passes the right parameters in specific situations. This is critical for actions like call transfers, data lookups, or external integrations.

Creating a Tool Call Test

Tool Call Testing Interface
1

Select the tool

Choose which tool you expect the agent to call in the given scenario (e.g., transfer_to_number, end_call, lookup_order).

2

Define expected parameters

Specify what data the agent should pass to the tool. You have three validation methods:

Exact Match
The parameter must exactly match your specified value.

Transfer number: +447771117777

Regex Pattern The parameter must match a specific pattern.

Order ID: ^ORD-[0-9]{8}$

LLM Evaluation An LLM evaluates if the parameter is semantically correct based on context.

Message: "Should be a polite message mentioning the connection"
3

Configure dynamic variables

When testing in development, use dynamic variable values that match those that would be actual values in production. Example: {{ customer_name }} or {{ order_id }}

4

Run and validate

Execute the test to ensure the agent calls the correct tool with proper parameters.

Critical Use Cases

Tool call testing is essential for high-stakes scenarios:

  • Emergency Transfers: Ensure medical emergencies always route to the correct number
  • Data Security: Verify sensitive information is never passed to unauthorized tools
  • Business Logic: Confirm order lookups use valid formats and authentication

Running Tests

Write tests for new behavior or known failures, run them while you iterate on prompts and configuration, then save once they pass.

Navigate to the Tests tab in your agent’s interface. From there, you can run individual tests, select multiple tests from your library as a batch, or execute your entire suite with Run All Tests.

Running tests on an agent

Probabilistic testing

Agent outputs can vary between runs. A single pass shows the agent can succeed; probabilistic testing shows how often it will by running the same test multiple times and reporting a pass rate.

Running a test multiple times

Split-run control on a test letting you pick how many times to execute it

When triggering a test from the dashboard, use the split-run control on the run button to choose how many times to execute it (for example 3×, 5×, or 15×). Each run is independent: the agent receives the same chat history, dynamic variables, and other inputs, but its response is generated fresh every time.

Multi-run works for individual tests, folders, and running the entire test suite attached to an agent. It’s compatible with all three test types — Simulation, Next Reply (Scenario), and Tool Call — and is typically most useful for Simulation tests, where the larger surface area of a multi-turn conversation makes response variation more likely.

Pass rates and result bucketing

Multi-run results grouped into pass and failure buckets with a pass rate badge

After a multi-run finishes, results are summarized as a pass rate (for example, 4/5 passed) with a colored badge:

  • Green — 100% passed
  • Amber — at least 80% passed
  • Red — below 80%

Individual runs are then grouped by failure reason so you can see how the agent fails, not just that it fails. Instead of scrolling through five separate transcripts to spot what differed, you see clusters like “Correctly routed to billing (4 runs)” and “Hallucinated a support number (1 run)”, each expandable to the underlying transcripts and evaluation rationale.

When to use it

  • Before shipping a change — Re-run attached tests probabilistically to confirm reliability hasn’t dropped (for example, from 95% to 60%).
  • Diagnosing flaky behavior — A single failure could be noise; a 1-in-5 failure with a clearly named failure bucket is a reproducible issue to fix.
  • Tuning prompts and tools — Iterate on configuration and compare pass rates side by side, rather than relying on one-off runs.

Running probabilistically via the API or SDK

Pass repeat_count (between 2 and 20) on the run-tests request to execute each test that many times. Setting repeat_count automatically enables failure bucketing on the response, so the returned invocation includes the per-bucket grouping and pass rate you’d see in the dashboard.

1from elevenlabs import ElevenLabs
2
3elevenlabs = ElevenLabs()
4
5invocation = elevenlabs.conversational_ai.agents.run_tests(
6    agent_id="<agent-id>",
7    tests=[{"test_id": "<test-id>"}],
8    repeat_count=5,
9)

Best Practices

Evaluate agent persona consistency

Test that your agent maintains its defined personality, tone, and behavioral boundaries across diverse conversation scenarios and emotional contexts.

Verify complex multi-turn reasoning

Create scenarios that test the agent’s ability to maintain context, follow conditional logic, and handle state transitions across extended conversations.

Test against prompt injection attempts

Evaluate how your agent responds to attempts to override its instructions or extract sensitive system information through adversarial inputs.

Assess ambiguous intent resolution

Test how effectively your agent clarifies vague requests, handles conflicting information, and navigates situations where user intent is unclear.

Next Steps