---
title: "Split Testing"
description: "A/B test agent variants on live traffic and compare per-variant outcomes."
---

Split testing routes an agent's live calls across **variants** —
weighted configurations that override specific fields (prompt, voice,
product, …) while inheriting everything else from the agent's deployed
config. Each call records which variant served it
(`agent_variant_id` / `agent_variant_label` on the
[Call object](/api-reference/calls#call-object)), and the stats
endpoint compares outcomes per variant.

The **control** variant (`is_control: true`) represents the agent's
deployed config unchanged. Enabling split testing auto-creates a
control if none exists.

## Endpoints

| Method | Path | Description |
|--------|------|-------------|
| `GET` | `/v1/agents/{agent_id}/variants` | List variants |
| `POST` | `/v1/agents/{agent_id}/variants` | Create a variant |
| `PATCH` | `/v1/agents/{agent_id}/variants/{variant_id}` | Update a variant |
| `DELETE` | `/v1/agents/{agent_id}/variants/{variant_id}` | Delete a variant |
| `POST` | `/v1/agents/{agent_id}/split-testing` | Enable / disable split testing (staged as a draft) |
| `GET` | `/v1/agents/{agent_id}/variant-stats` | Per-variant call stats |

## Variant object

```json
{
  "id": 41,
  "label": "Warmer greeting",
  "weight": 50,
  "is_control": false,
  "prompt": "You are a warm, upbeat support agent…",
  "voice": null,
  "product": null,
  "background_track": null,
  "thinking_level": null,
  "acknowledgement_prompt_mode": null,
  "acknowledgement_prompt": null,
  "inbound_speak_order": null,
  "outbound_speak_order": null,
  "silence_interval_seconds": null,
  "silence_max_checkins": null,
  "created_at": "2026-04-20T18:24:10.113Z",
  "updated_at": "2026-04-20T18:24:10.113Z"
}
```

| Field | Type | Description |
|-------|------|-------------|
| `id` | integer | Variant id |
| `label` | string | ≤ 100 chars |
| `weight` | integer | 0–100, default 50. Traffic is distributed proportionally to weights |
| `is_control` | boolean | At most one control per agent |
| `prompt`, `voice`, `product`, `background_track`, `thinking_level`, `acknowledgement_prompt_mode`, `acknowledgement_prompt`, `inbound_speak_order`, `outbound_speak_order`, `silence_interval_seconds`, `silence_max_checkins` | field type \| null | **Overrides.** `null` = inherit the agent's deployed value; set a value to change just that field for this variant |

`POST` requires `label` (everything else optional) and returns `201`.
`PATCH` accepts any subset; `DELETE` returns `204`.

---

## Enable / disable split testing

<CodeGroup>
```bash cURL
curl -X POST https://api.thunderphone.com/v1/agents/12/split-testing \
  -H "Authorization: Bearer sk_live_YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"enabled": true}'
```
</CodeGroup>

| Field | Type | Default | Description |
|-------|------|---------|-------------|
| `enabled` | boolean | `true` | |

```json Response
{
  "split_testing_enabled": false,
  "draft_split_testing_enabled": true
}
```

The toggle is staged as a **draft** on the agent — live traffic starts
(or stops) splitting only after you
[deploy](/api-reference/agents#deploy-the-draft). Enabling with no
control variant auto-creates one (`label: "Control"`, `weight: 100`).

---

## Variant stats

<CodeGroup>
```bash cURL
curl https://api.thunderphone.com/v1/agents/12/variant-stats \
  -H "Authorization: Bearer sk_live_YOUR_API_KEY"
```
</CodeGroup>

```json Response
[
  {
    "variant_id": 40,
    "label": "Control",
    "is_control": true,
    "calls": 132,
    "success_rate": 84.1,
    "avg_duration_seconds": 163
  },
  {
    "variant_id": 41,
    "label": "Warmer greeting",
    "is_control": false,
    "calls": 129,
    "success_rate": 88.4,
    "avg_duration_seconds": 171
  }
]
```

| Field | Type | Description |
|-------|------|-------------|
| `calls` | integer | Total calls served by the variant |
| `success_rate` | number \| null | % of graded, completed calls with `call_outcome="success"` (ungraded calls excluded); `null` when nothing graded yet |
| `avg_duration_seconds` | integer \| null | Mean duration of completed calls |

---

## Related

<CardGroup cols={2}>
  <Card title="Agents" icon="robot" href="/api-reference/agents">
    `split_testing_enabled` lives on the agent (draft/deploy).
  </Card>
  <Card title="Calls" icon="phone" href="/api-reference/calls">
    Filter call history by `agent_variant_id` fields on each call.
  </Card>
</CardGroup>