Evals and Test Center

Test a voice agent before it goes live. Use the Test Center for browser calls, text chat and variable testing, and eval suites with assertions, simulated conversations and an LLM judge.

3 min readUpdated

Edesy gives you two ways to test an agent. The Test Center is for trying the agent yourself: a browser voice call, a text chat, a phone test call and a variable tester. Evals are repeatable test suites: each case has a scenario and assertions, and runs as a simulated text conversation graded by rules and, optionally, an LLM judge.

Test Center#

Mode What it does Uses credits
Browser call Talk to the agent from your browser over WebRTC Yes
Text chat Chat with the agent's logic in text, without voice Model usage only
Phone call The agent calls your phone Yes
Variable tester Fill {{variables}} (for example customer_name) before a test -
Scenarios Reuse saved test setups -

Open any agent and click Test. After a test call, the call appears in call history with its transcript, tool calls and latency breakdown.

Evals#

An eval suite belongs to one agent and holds cases. Each case has:

  • a scenario: variables, the caller's lines, and optional mocked tool results
  • assertions: what must be true of the conversation
  • a severity: low, medium, high or critical
  • optional tags

Assertion types#

Assertion Passes when
assistant_contains The agent says a given text
assistant_not_contains The agent never says a given text
tool_called The agent calls a given tool
tool_not_called The agent never calls a given tool
tool_argument_equals A tool is called with a given argument value
minimum_turns The conversation has at least N turns
maximum_latency_ms The agent responds within your limit
llm_rubric An LLM judge decides the conversation meets your rubric

A case where every assertion was skipped is reported as Skipped, not as passed.

How a run works#

Each run is a simulated text conversation. The case's caller lines drive the conversation, and mocked tool results stand in for your real APIs. Rule-based assertions are graded straight away; llm_rubric assertions are graded by the LLM judge.

Evals run in text. Voice-mode evals are not available.

Run a suite#

  1. Open the agent and go to Evals.
  2. Create a suite, then add cases with a scenario and assertions.
  3. Click Run. Each run records the agent configuration and prompt version it tested.
  4. Open the results to see each assertion's evidence, the transcript and the tool trace.
  5. Change the prompt, run again and compare.

Example case#

{
  "name": "Asks for order number before lookup",
  "severity": "HIGH",
  "scenario": {
    "variables": { "customer_name": "Asha" },
    "userTurns": ["Where is my order?", "It is 45821"]
  },
  "assertions": [
    { "type": "tool_called", "tool": "lookup_order" },
    { "type": "tool_argument_equals", "tool": "lookup_order", "path": "order_id", "value": "45821" },
    { "type": "llm_rubric", "rubric": "The agent stays polite and does not invent a delivery date." }
  ]
}

FAQs#

Do evals place real calls? No. They run as text conversations, so they are quick to repeat after each prompt change.

Can evals score live calls automatically? No. Automatic QA of live calls is not available. Use extraction for structured post-call data.

Can I test tools without hitting my real API? Yes. Add mocked tool results to the scenario.

What should a first suite cover? The happy path, a caller who refuses, a caller who asks something off-topic, and each tool your agent uses.

Was this page helpful?