Skip to main content

​
What is Personality in Evaluators?

A personality defines how your evaluator behaves during conversations with your AI agent. It determines the language, tone, and behavioral characteristics of the testing agent, allowing you to simulate diverse real-world user interactions.

​
Personality vs. evaluator

The two answer different questions, and mixing them up is the most common source of tests that quietly do not test what you meant:
Evaluator instructions cannot override a personality setting. “Speak with a strong accent”, “interrupt constantly”, “talk very quietly” and “stay silent for 30 seconds” have no effect when written into instructions — the call runs and the result comes back, but the behavior never happened. Anything that lasts the whole call belongs on the personality.
The rule of thumb is duration. A trait that colours every turn (an accent, a speaking pace, ambient noise, a poor connection) is a personality. A single moment (“gets frustrated at step 4”, one interruption) belongs in the evaluator’s instructions.

​
Why Use Different Personalities?

Testing your AI agent against different personalities helps you:
  • Identify Edge Cases: Discover how your agent responds to challenging conversational patterns
  • Test Real-World Conditions: Simulate environments with background noise or interruptions
  • Ensure Accessibility: Verify your agent works well with different languages and speaking styles

​
Personality Components

​
Language

The primary language in which the evaluator will communicate with your agent. This ensures your AI agent is tested in the languages it’s designed to support. Choose it first — personalities are language-specific, and an evaluator’s scenario_language must match its personality’s language.

​
Voice (and therefore accent)

The voice is what the caller actually sounds like, and it is where the accent comes from. The accent shown on a personality is a read-only label derived from its voice, not a setting: speech is synthesised from the voice ID alone, so testing a different accent means choosing a different voice. Browse the catalog with GET /test_framework/v1/personalities/elevenlabs_voices/ (which publishes an accent label per voice) or GET /test_framework/v1/personalities/cartesia_voices/, then set that id on the personality. If no listed voice has the accent or language you need, Cekura can add voices on request — contact support@cekura.ai or your dedicated Slack support channel.

​
Behavioral Characteristics

Personalities control several behavioral aspects:
  • Background Noise: Simulates real-world conditions like calling from a busy street or office. To use a custom audio file as background — including recordings with actual voices — select the Custom URL option when creating a personality and provide a direct link to any audio file (e.g. https://your-domain.com/background.mp3). See Creating Custom Personalities for setup steps.
  • Interruption Patterns: Tests how your agent handles being interrupted mid-sentence
  • Speaking Pace: Varies from slow and deliberate to fast and hurried
  • Emotional Tone: Ranges from calm and patient to frustrated or urgent
  • Idle (stall) Behavior: How long the testing agent waits in silence before prompting (“Are you still there?”), and how many times it prompts before giving up. Defaults are 10 seconds and 3 prompts.
Behavior in this list is set on the personality and cannot be overridden from an evaluator. Instructions, expected outcomes, and your agent’s description control what the testing agent says — not how long it stays silent, how fast it speaks, or when it interrupts.If a test needs the testing agent to stay quiet longer than the idle timeout, raise Idle Timeout on the personality. Because the pre-defined personalities are shared across all workspaces, they can’t be edited in place — fork one into your project first, then change the timeout on the fork. See Creating Custom Personalities.For a pause in one specific step rather than call-wide, use a <hold> tag in a conditional action instead. The idle timer is paused for the duration of a hold, so a hold longer than the idle timeout does not trigger the idle prompt — you only need to change the personality when the silence is not inside a hold.

​
Example Personalities

​
Assigning Personalities to Evaluators

When creating or editing an evaluator, you can select from:
  1. Pre-defined Personalities: Cekura’s curated collection of common user types
  2. Custom Personalities: Your own created personalities for specific testing needs
Each evaluator can have one personality assigned to it, ensuring consistent behavior across multiple test runs.

​
Creating Custom Personalities

For detailed instructions on creating your own personalities, see Creating Custom Personalities.

​
Best Practices

Test with Multiple Personalities: When running evaluators, you can select multiple personalities to ensure your agent performs consistently across user types.
For comprehensive testing, consider creating evaluators with:
  • 60% standard, clear communication
  • 20% challenging conditions (background noise, interruptions)
  • 10% non-native speakers or accent variations
  • 10% edge cases (frustrated users, very fast/slow speakers)
This distribution mirrors real-world user demographics while ensuring thorough testing coverage.