A queue of 100 Shorts: unique input and a cap per item

Bulk runs let you leave a list of Format runs to work overnight. The Sume docs, read on October 5, 2026, say a bulk request is a server-side queue of ordinary Format runs: each item is one sandbox, one agent turn and one run receipt. The limits are 1 to 100 items and a concurrency window of 1 to 16.

The danger is sameness, not volume

YouTube's monetization page describes repetitive or mass-produced AI content from generic templates. A queue of 100 items is not a violation by itself. A queue of 100 items that differ only by a product name is the risk. Treat the queue as a way to run 100 different ideas, not 100 copies of one idea.

What each item carries

Each entry in items has the same body as a single run, and must name at least one of instruction, input, previous_run_id or attachments. Each item can bind its own output schema and its own generation_spend_cap_usd, and its own webhook, according to the bulk runs page.

That per-item cap is your safety net. One runaway item cannot spend the whole batch.

Make the inputs differ on purpose

Before you submit, read the 100 instructions as a list. Count distinct opening lines, distinct lengths and distinct proof sources. If fewer than most are distinct, rewrite the list, not the Format.

When the queue finishes, branch on counts.failed. The docs say completed means every item is terminal, not that every item succeeded. Mint a fresh Idempotency-Key per batch, since replaying a spent key returns the old queue.

Budget the batch

Spend caps are set in dollars per item, so do the arithmetic before you submit. If the cap is a few dollars per item and you queue 100, the worst case is 100 times that cap, even though most items will cost far less. Pick a cap you would be happy to pay on every item, and run a small batch of five first to see the real figure from the receipts.

Then review the finished Shorts as a set. A human glance at the first frame of each is a cheap similarity check, and it takes a few minutes for 100 thumbnails laid out in a grid. Remove or redo any that look like siblings.