Sume Agent Completions vs Format vs Scheduled runs: request body diff

Sume's Format runs, Scheduled (Action) runs, and Agent Completions run the same agent and return the same receipt shape, but their request bodies are not interchangeable: the instruction source, the spend cap's default and null, on_active_run, attachments, unknown-field handling, scopes, and webhook events all differ by surface.

The comparison comes from Sume's Create a run, Advanced: run a schedule via API, and Agent Completions docs pages, read on 2026-09-26. For which surface fits which job, see What is a video agent?

Which request fields differ between the three surfaces?

Scheduled is the product name; the wire namespace is /v1/actions. Each surface differs in where the instruction comes from and what Sume has saved for you, and the body follows from that.

What does null mean for the spend cap on each surface?

The field has one name and three rule sets. On a Format run, a number up to 500 is honored even above the Format's own cap, null runs at the $500 platform maximum, and 0 or anything above 500 is a 400. On a Scheduled run, a number is clamped to the lower of the request and the schedule's cap, so it can lower the cap but never raise it; null runs with no automation ceiling, while wallet balance, generation admission, and org limits still apply; 0 is a 400. On an Agent Completion the cap is required and has no default, and the Agent Completions page does not document a null value. More in spend caps for unattended AI agents.

Why does a body copied between surfaces misbehave?

Three rules catch people who port a body from one surface to another:

• on_active_run: a Format defaults to allow and runs concurrently, subject to workspace generation concurrency, while a schedule defaults to skip. The Format docs say not to copy schedule bodies into Format calls. See how to prevent overlapping agent runs.
• Unknown fields: a Format answers 400 unknown_parameter, with a suggestion when the name is close (webook_url → webhook_url). A schedule accepts only its listed properties and silently drops unknown top-level ones, so a typo fails quietly.
• Size: a Format body is capped at 4 MiB (413 payload_too_large). input is capped at 64 properties and 2 MiB on both Formats and schedules.

How do idempotency and webhooks compare?

On all three, replaying an Idempotency-Key with the same payload returns the original receipt with idempotency_hit: true, and reusing it with a different payload is 409 idempotency_conflict. Format keys are scoped to one Format and run up to 255 characters. On a schedule, a request without a key has no replay protection, so every such request starts a new run.

communication.webhook_url works the same on all three surfaces: one signed POST when the run completes or fails, carrying the same receipt the poll endpoints return. The event name tells you the family. A skipped run and a canceled run never deliver one.

Which run id works where?

Each family lives behind its own URL prefix: /v1/format-runs/…, /v1/action-runs/…, and /v1/agent-runs/…. Format and Action run ids both start with arun_, so the id cannot tell those two apart: store the surface next to the id. An Action or Format run id does not resolve on /v1/agent-runs (404 agent_run_not_found). In the TypeScript SDK, waitForRun requires family: "format", "action", or "agent".

Service-account keys cannot create Format runs, Scheduled runs, or Agent Completions. Each refusal is 403 insufficient_scope, with a details.reason naming the surface.