Testing Conversational AI Agents

When deploying a conversational voice agent into real-world services, the key challenges are how to monitor the agent's behavior at scale, verify it works as intended, and test changes.

ElevenLabs, while developing El, a documentation support agent, built a system for monitoring, evaluating, and testing agents based on evaluation criteria and conversation simulation.

To measure the agent's performance, they defined the following Evaluation Criteria:

• Interaction: validity of the conversation and relevance of the questions
• Positive interaction: user satisfaction and whether confusion/frustration occurred
• Understanding the root cause: whether the user's underlying issue was identified
• Solving the user's enquiry: whether the issue was resolved or an alternative was offered
• Hallucination: whether information not present in the knowledge base was generated

For the testing phase, they use the Conversation Simulation API. This API simulates real user scenarios and supports two approaches:

• Full simulations: testing the entire conversation from start to finish
• Partial simulations: a unit-test-style approach that starts mid-conversation to verify specific decision points or sub-flows

Finally, they integrated evaluations and simulations into their GitHub DevOps workflow to implement automation within the CI/CD pipeline. Every update is automatically tested before deployment, preventing regressions and providing fast feedback on real-time performance.