Comfy Agent Ask or Auto mode vs unattended Sume Format runs

Comfy Agent has two run modes: Ask, which waits for your approval before each workflow run, and Auto, which runs without asking. A Sume Format run over the API has neither. It is unattended by definition: approvals are treated as granted, and the controls that matter move to the request, namely a spend cap per run, a key scope, and what the run does when it cannot finish.

If you are picking a tool for an overnight batch, the useful question is not which one asks. It is what limits the damage once nobody is there to ask.

How do the two approval models compare?

The Comfy facts come from its in-app Agent page; the Sume facts come from the Format API docs.

What does Ask mode actually buy you?

Ask mode puts a person between a plan and its cost, one run at a time. It suits graph building, where you are watching the canvas anyway and a wrong node costs a credit or two. It stops being useful the moment you want fifty variations while you sleep, because fifty approvals is a babysitting job, and Auto removes the gate without replacing it. The Agent page does not describe a per-run spending cap, so with Auto the pool of credits is the only limit it mentions.

What replaces the prompt in an unattended Sume run?

Four things in the request, all documented on Calling a Format and Errors and spend.

• A spend cap on every run. generation_spend_cap_usd takes a number above 0 and up to 500; omit it and the Format's own cap applies. Compare usage.billable_amount_usd_micros with the cap on the receipt to see how close a run came.
• A key that can only do what the job needs. Creating runs and queues needs formats:write, and reading them needs formats:read. Scopes are fixed when a key is minted.
• A refusal to guess. A run that hits a gate only a person could pass, such as a missing input, ends failed with unattended_blocked and a message written to be shown, rather than quietly delivering something partial.
• An overlap rule. on_active_run set to skip or reject keeps a second request from starting while one is in flight. In a bulk queue every item is forced to allow, so use the window of 1 to 16 as your throttle instead.

How do I keep a person in the loop on Sume anyway?

Run a pilot. Send one item as an ordinary run, look at its output, and only then send the other ninety-nine as a bulk queue. That puts the approval in front of the batch rather than in front of each row, which is the trade Ask mode makes at a finer grain.

This sends the pilot with a tight cap:

curl -sS -X POST https://api.sume.com/v1/formats/myteam/product-promo/runs \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: pilot-autumn-001" \
  -d '{"instruction":"Pilot clip for approval.",
       "input":{"sku":"A-100"},
       "generation_spend_cap_usd":15}'

Which should you choose?

Choose Comfy Agent when the work is building and tuning a workflow with your eyes on the canvas, and pick Ask mode while you learn what each run costs. Choose Sume when the output is a finished clip from a saved brief, repeated across rows. Sume does not give you a node graph to edit; a Format is a saved brief, so changing the approach means editing the Format and not rewiring nodes. Comfy's page lists local Agent use as coming soon, so a self-hosted pipeline still has to use Comfy's separate MCP or CLI tools for now.