script_run or Agent Completion: which for a GPT-6.1 Sol agent?

Use script_run when your coding agent already knows the calls: three or more independent tool calls of the same shape, such as one image per scene. Use an Agent Completion when the task itself is open-ended and Sume's own agent should decide the steps. script_run runs in a bounded 5 to 55 second window on the MCP session; a completion is an asynchronous run with its own sandbox, started over the API with a required spend cap.

What each one is

MCP tools and gates describes script_run as a short JavaScript program run on the Sume side that calls listed tools in a loop, in parallel or conditionally, and returns one value. It is bounded by timeout_seconds (5 to 55), max_calls and max_paid_calls, and paid creates inside it still need their own idempotency_key. The Agent Completions page describes an unattended run of the Sume Agent with a sandbox, tools, an MCP bridge and media generation, answering 202 with a receipt you poll.

Side by side

Rules of thumb

• Fan-out of identical, well-specified calls: script_run. Six scene stills, ten caption passes.
• Read-heavy discovery that needs judgment: let the coding agent call read tools directly.
• A brief like "make a teaser from this URL" where the steps are unknown: Agent Completion.
• Anything that must survive the coding session ending: Agent Completion, or save job ids from script_run and wait later.

Limits to remember

Discovery tools and script_run itself cannot be called from a script. script_run returns child jobs[] you still jobs_wait on, one call holding up to 55 seconds. A completion has no streaming and no thread continuation yet, and assistant turns in messages[] are rejected. If you need a saved, repeatable workflow with changing inputs, a Format is a third option, and Scheduled runs saved instructions on a clock.

Whichever route you use, the model on your side does not have to be GPT-6.1 Sol. The same split applies to any coding agent that can speak MCP or HTTP.