ElevenLabs provides a dedicated testing framework for conversational voice agents. Unlike traditional software tests that validate deterministic outputs, voice agent testing must account for the probabilistic nature of LLM responses, correct tool invocation under varying inputs, and the branching structure of multi-turn dialogue. The platform addresses each of these with distinct test types and a simulation engine.

Tests can be run manually, triggered via API, or integrated into CI/CD pipelines so that every change to an agent's prompt or workflow is validated before it reaches production. This page covers the six core testing concepts: scenario evaluation, tool call validation, conversation simulation, CI/CD integration, test generation from real conversations, and the API surface.

Tool call tests verify that the agent invokes the correct tool with the correct parameters given a particular conversational context. This is distinct from scenario testing: instead of evaluating whether the agent's spoken response is correct, it validates the function call the agent would make before generating a response. For actions like transferring a call, looking up an account, or booking an appointment, the parameter values are not merely important — they are the outcome.

Voice agent testing integrates into standard CI/CD pipelines via the REST API or the provided SDKs. Every change to the agent's system prompt, tool definitions, or conversation workflow can trigger a test suite run automatically before deployment. This prevents regressions from reaching production, which matters especially for voice agents where a degraded interaction is immediately perceived by callers.

- name: Run ElevenLabs agent tests
  env:
    ELEVENLABS_API_KEY: ${{ secrets.ELEVENLABS_API_KEY }}
    AGENT_ID: ${{ vars.AGENT_ID }}
  run: |
    python scripts/run_agent_tests.py \
      --agent-id "$AGENT_ID" \
      --fail-threshold 0.95 \
      --timeout 300

The full testing lifecycle is accessible via the REST API. Test suites can be created, updated, executed, and their results retrieved programmatically. Both the Python and TypeScript SDKs wrap these endpoints for typed, idiomatic usage in each language.

from elevenlabs import ElevenLabs
import time

client = ElevenLabs(api_key="your-api-key")

# Trigger test suite
run = client.convai.agents.run_tests(
    agent_id="your-agent-id"
)

# Poll until complete
while run.status in ("pending", "running"):
    time.sleep(5)
    run = client.convai.agents.get_test_run(
        agent_id="your-agent-id",
        run_id=run.id
    )

# Evaluate results
passed = [t for t in run.tests if t.passed]
failed = [t for t in run.tests if not t.passed]
print(f"{len(passed)}/{len(run.tests)} tests passed")

import { ElevenLabsClient } from "@elevenlabs/elevenlabs-js";

const client = new ElevenLabsClient({
  apiKey: "your-api-key",
});

// Trigger test suite
let run = await client.convai.agents.runTests({
  agentId: "your-agent-id",
});

// Poll until complete
while (run.status === "pending" || run.status === "running") {
  await new Promise(r => setTimeout(r, 5000));
  run = await client.convai.agents.getTestRun({
    agentId: "your-agent-id",
    runId: run.id,
  });
}

// Evaluate results
const passed = run.tests.filter(t => t.passed);
console.log(`${passed.length}/${run.tests.length} passed`);