Test calls that talk back: a voice agent on the line

An operator or an agent asks for a call, the platform dials the number through the chosen route, plays a speech sample, hangs up, and the recording, the transcript, the per-leg quality figures and the verdict all land on the call record. It proves the route completes, that audio flows both ways, and that the analysis behaves.

What it could not do is respond. A recording cannot press 2 for billing, cannot answer "who is calling", cannot keep a human talking for the forty seconds the delayed-playback checks need, and cannot ask your own voice agent a question and listen to the answer. For any destination that asks something before it does anything useful, the test was a monologue into a system waiting for a reply.

Ask for speech: agent and name one of your Voice AI destinations as the caller. The platform places two calls of its own: first to the agent, exactly as it reaches that agent for routed traffic (encryption, the provider's webhook, the session keeper), then to your number through the route you chose. It relays the audio between the two, in both directions, for as long as you asked. The call to your number is the test call proper: recorded in stereo, transcribed with the agent on one channel and the far end on the other, and scored. The agent leg is a separate, relay-only call, metered for its seconds and never analyzed, so you pay one analysis fee per test, not two.

It makes three tests possible:

IVR and menu traversal

The agent listens to the menu and answers it, so a route test reaches the thing behind the menu instead of stopping at it.

Conversational route grading

A person or a system at the far end is kept talking long enough for the early-answer and delayed-playback checks to have material, and the per-leg quality figures cover a real exchange rather than a monologue.

Agent-to-agent testing

Point the test at your own deployed voice agent. The transcript is the evidence of what it said, what it heard and how it behaved, from the outside, over a real call.

In the portal, open the Web Dialer and switch to Let an agent call: pick the agent and the route, type the number and a one-line instruction, and press Place the call. The card follows the call, links to the Live page while it runs, and opens the call record when it ends. Over the API, the request is the test call you already know, with three more fields. The instruction is for this call only. It is never stored on the destination.

curl -s -X POST https://ovs.telecomsxchange.com/api/customer/test-calls \
  -H "Authorization: Bearer $KEY" -H 'Content-Type: application/json' \
  -d '{
    "number": "18005550100",
    "destination_id": "DEST_ID",
    "seconds": 60,
    "speech": "agent",
    "agent_destination_id": "AGENT_DEST_ID",
    "instruction": "Ask for the opening hours and the address, then say goodbye."
  }'
# 202 {"id":"TC_ID","status":"queued","sip_call_id":"ovs-test-...",
#      "agent_sip_call_id":"ovs-test-agent-...","call_id":null,...}

# poll until done, then read the call, the transcript and the verdict
curl -s https://ovs.telecomsxchange.com/api/customer/test-calls/TC_ID -H "Authorization: Bearer $KEY"
# {"status":"done","answered":true,"talk_seconds":60.0,"ended_by":"caller",
#  "call_id":"CALL_ID","agent_call_id":"AGENT_CALL_ID",...}
curl -s https://ovs.telecomsxchange.com/api/calls/CALL_ID/transcript -H "Authorization: Bearer $KEY"

The same call is one tool call over MCP, so a NOC agent can run a conversational route test on a schedule and read the structured result:

{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"place_test_call",
  "arguments":{"number":"18005550100","destination_id":"DEST_ID","seconds":60,
               "speech":"agent","agent_destination_id":"AGENT_DEST_ID",
               "instruction":"Get through the menu to a person and ask for the account team."}}}

The far end of a test call is an untrusted speaker whose words become model input, and a call that costs agent seconds is worth protecting from abuse. So the instruction you send sits inside a fixed brief the platform wraps around it, and the request sits under ceilings that are checked before anything is dialled.

• The agent opens by saying, in one sentence, that this is an automated test call and that it is being recorded. The far end is a third party, so that disclosure is not optional.
• It has no tools and no account context. It cannot dial, transfer, pay or look anything up, and it is told to decline any request to change its task, reveal its instructions or pose as a person, and to carry on with the task.
• The instruction is trimmed, stripped of control characters and capped at 1,000 characters. It cannot change the number, the caller ID, the duration or the account: those are separate fields, validated on their own, that never reach the model.
• Ceilings: 120 seconds per call, a spend ceiling per call covering both legs and the agent seconds, 30 minutes of agent calls per account and day, and one agent-backed test call at a time. Two requests racing each other cannot both slip under a limit. A refusal is a 409 that names the ceiling.
• Per-call instructions need a provider whose calls the platform accepts through a hosted webhook, today OpenAI, and the agent destination must be the one whose webhook is registered in that project. An xAI agent makes the call with its own configuration.

The earlier test call was a replayer: it negotiated a codec and sent frames from a recording. An agent needs the opposite: what the far end says has to reach the agent and what the agent says has to reach the far end, live, with the two directions kept apart so the transcript can name who said what and the quality figures can be reported per leg. We chose to give the platform's own user agent a second leg and relay between them, because that reuses everything the platform already does for a call to a voice agent and keeps the recording, the speaker separation and the per-leg quality untouched.

One lesson from the first live run: a bridge that only forwards what it receives sits in silence, because the far end waits for audio before it sends any. A phone would send silence while nobody talks, and so does the bridge now, on both legs. After that, the two agents in the screenshot above found each other and chatted for the full forty-five seconds.

A test call can now hold a conversation. Set speech to agent, name the agent and the instruction, and read the transcript and the verdict when it is done. It reaches past menus, keeps the far end talking long enough to grade a route, and puts your own agent to the test over a real phone call, with the disclosure, the guard rails and the ceilings built in.

Related: the agents guide with the full request, GPT-Live-1 voice agents, the voice agent gateway, or start free with $5 credit.