TikTok Symphony Agent makes 3 variations at a time: run three cuts

TikTok's Symphony Agent returns up to three video variations at a time for A/B testing, and it can generate a 15-second video in about five minutes (read 2026-10-03). If you want the same three-way test from a master you already own, Sume cannot write the ad, but it can cut three versions in parallel: three trim requests with different starts, lengths or caption placements, each with its own idempotency key. That gives you the testing set without waiting for an agent.

This post covers what TikTok says Symphony Agent does and a script for three cuts.

What TikTok's page says

The page describes an AI-powered assistant inside Symphony Creative Studio. It accepts product images, selling points, target audience, brand tone and creative constraints, and it can start from an ad trend, a TikTok template, your own reference video or a free-text prompt. Outputs are a product brief, an insight report, a storyboard and a finished video. Access is based on the ad spend of the TikTok Ads Manager account linked to Symphony Creative Studio, the tool is desktop only, human faces in product images are flagged for review, and non-compliant images and text are blocked before generation (read 2026-10-03).

Three cuts from one master

A useful test changes one thing at a time. With video trim the knobs are start, duration, and optionally an output size and audio. A hook test, for example, can use three different start times from the same clip, each with the same duration, so the only difference is the opening seconds. Add a video captions job on top if you want to test text placement through design.placement.anchor_ratio.

The script submits the three trims, each with its own idempotency key, then polls them. Trim costs $0.02 per job under the current public rate, so the set is three jobs (confirm in GET /v1/catalog). A webhook is available in the docs if you would rather not poll.

import os, time, requests
BASE = "https://api.sume.com"
H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def run(path, body, key):
    r = requests.post(BASE + path, headers={**H, "Idempotency-Key": key}, json=body)
    r.raise_for_status()
    job = r.json()["request_id"]
    while True:
        s = requests.get(f"{BASE}/v1/jobs/{job}/status", headers=H).json()["status"]
        if s in ("completed", "failed", "canceled"):
            break
        time.sleep(3)
    return s, (requests.get(f"{BASE}/v1/jobs/{job}/result", headers=H).json() if s == "completed" else None)

MASTER = "https://media.sume.com/artifacts/artf_demo/master.mp4"
for name, start in [("a", 0), ("b", 4), ("c", 9)]:
    status, res = run("/v1/video-trim", {
        "video_url": MASTER, "start": start, "duration": 15}, f"hook-test-{name}-001")
    print(name, start, status, res and res["video_url"])

Designing the three-way test

Decide the question before you cut. If you ask which opening works best, keep length, captions and music identical and change only the start. If you ask whether shorter is better, keep the start the same and set duration to 9, 15 and 21 seconds. Changing both at once makes the result hard to read, because you cannot tell which change moved the number.

Keep the sample fair. Give each variation the same audience, budget and run time inside Ads Manager, and let the test run long enough to collect a meaningful number of views before you pick a winner. TikTok's page describes the A/B testing use for Symphony outputs but does not give a sample size, and neither does Sume, so set the rule yourself before you look at the numbers.

Finally, label the files. A name such as hook-b-start4-15s.mp4 is easier to match to a report than a job id, and it lets a teammate repeat the test next month. Store the idempotency key in the same note, so a retry of one cut is not mistaken for a new variation.

Be clear about what differs

Symphony Agent writes new scripts and visuals, and Sume's trim only rearranges material you already have. Three cuts from one master will share the same story and the same footage, so use them to test openings, lengths and captions, not offers or messages. If you want three different concepts, you need three different masters.

Record the three versions with a note on what changed, and keep the master. When results come in, the notes tell you whether the winning cut won on its opening, its length or its captions.