The agentic UGC pipeline: generate, post, measure, repeat

The loop

1. Generate. The agent, or a generator it calls, produces a batch: hooks, slideshows, avatar videos, captions.
2. Post. The batch is spread across accounts on a cadence with POST /v1/campaigns (or create_campaign over MCP).
3. Measure. Views, likes, comments, shares and saves come back from TikTok and Instagram every hour.
4. Decide. The agent ranks what worked, retires angles that didn't, and writes the brief for the next batch.

UGC API handles steps 2 and 3. It doesn't generate content; you bring your own generator or model. That split is deliberate: the posting layer stays boring and reliable while you swap models freely.

Step 1: generate with a test in mind

An agent learns faster when each batch tests one thing. Ask it for 12 hooks on the same product and the same visual, varying only the first line. Or the same hook in three formats: a talking-head video, a 5-slide slideshow and a single image. Give every piece a short label like hook=price-shock and keep it in your own database with the post id you get back. That label is what turns numbers into decisions.

Step 2: post across accounts

Spread the batch so no account gets flooded and each angle runs on more than one account. A campaign does this in one request:

POST /v1/campaigns
{
  "name": "desk-hooks-wk44",
  "accounts": ["acc_tt_a", "acc_tt_b", "acc_tt_c", "acc_ig_a"],
  "posts": [
    { "format": "video", "media": ["https://cdn.example.com/h01.mp4"], "caption": "I wasted $400 on desk gear so you don't have to" },
    { "format": "slideshow", "media": ["https://cdn.example.com/s1.jpg", "https://cdn.example.com/s2.jpg", "https://cdn.example.com/s3.jpg"], "caption": "3 desk upgrades under $30" }
  ],
  "cadence": { "per_day": 2 },
  "times": ["10:00", "19:00"],
  "timezone": "America/New_York",
  "start_date": "2026-11-02"
}

Posts are assigned round-robin and staggered by a few minutes. In a chat assistant, the same thing is preview_campaign then create_campaign, and the preview shows every slot before anything is scheduled. See campaigns and UGC posting cadence.

Step 3: measure

UGC API reads engagement for published posts every hour: Instagram numbers from Instagram insights, TikTok numbers from TikTok's video list. Two calls cover what an agent needs:

• GET /v1/stats?from=…&to=… returns results by status, platform and day, totals for views, likes, comments, shares, saves and reach, and the five top posts by views. Over MCP it's get_stats.
• GET /v1/posts/:id returns each result with its own metrics and metrics_at, so you can join numbers back to your labels. GET /v1/posts?campaign_id=… lists a whole campaign.

Give posts time before judging. A TikTok can pick up views days after posting. A fair rule: compare posts at the same age (say 48 hours), and only across the same account or similar accounts. Details in stats.

Step 4: decide

Hand the agent a compact table, not raw JSON: label, platform, account, age, views, likes, shares, saves. Then ask specific questions:

Shares and saves are often better signals than likes for UGC that sells. Ask the agent to rank by a simple score you choose, like views + 20×shares + 10×saves, and keep that formula fixed so batches compare fairly.

A minimal implementation

The whole loop fits in one scheduled job and one table. Each night the job: reads yesterday's results from GET /v1/stats and GET /v1/posts, joins them to your labels, asks the model for next week's brief and batch, generates the content, and creates a campaign. Store per piece: your label, the media URL, the post id, and the metrics you read back.

pieces
  id            text   -- your id
  label         text   -- hook=price-shock, format=slideshow
  media_url     text
  post_id       text   -- from POST /v1/campaigns or /v1/posts
  account_id    text
  platform      text
  published_at  timestamp
  views_48h     int
  shares_48h    int
  saves_48h     int

Keep the model's role narrow: rank labels, propose the next tests, write captions. Keep scheduling rules (cadence, times, accounts, budget) in code, where they can't drift from one run to the next.

Where people stay in the loop

• Approve the brief, not every post. A person signs off on the angles each week; the agent produces and schedules inside them.
• Look at upcoming posts. Schedule at least a few hours ahead and skim GET /v1/posts?sort=scheduled_at. Cancelling is free: unsent posts are refunded.
• Watch failures the same day. post.failed and account.disconnected webhooks catch a broken account before it eats a week of slots.
• Use drafts for sensitive brands. TikTok drafts let a creator review and add a sound before posting.

Mistakes that break the loop

• Changing too much at once. If the hook, the visual and the account all change, the numbers can't tell you which mattered.
• Judging too early. A post that's 6 hours old isn't comparable to one that's 3 days old. Compare at a fixed age.
• Letting the agent pick the cadence. Posting rates belong in code. A model that sees a winner may want to post it ten times today.
• Ignoring failures. A disconnected account shows as zero views, which looks like bad content. Filter out failed and unpublished results before ranking.

Running it from a chat assistant

You don't need to write the loop to start. Connect Claude or ChatGPT to the MCP server and run it by hand once a week: "Here are 12 new videos, spread them over my 4 accounts from Monday", then a week later, "Which of last week's posts did best by shares, and what do they have in common?" When the weekly ritual is stable, move it into code with an agent framework. See agent frameworks.

What a month costs

Ten accounts posting twice a day is about 600 posts a month: $6 at $10 per 1,000. Metrics reads are free. Your model and generation costs will be the bigger line, which is the point: posting shouldn't be what limits how many tests you run.