Human in the loop AI agents: where a person approves

Human in the loop means an AI agent stops for a person before a step that matters, such as spending money, publishing, or deleting, and waits for approval or a correction. An agent without a person in the loop acts on its own, so any limits have to be set in advance. Sume's video agent works both ways: in chat it asks before it spends, while API and scheduled runs are unattended, so a spend cap stands in for the approval and a run that truly needs a person comes back failed instead of half-finished.

Sume's side comes from its Quick start, Format API, Create a run, and Agent Completions docs, read on 2026-09-27, and from current server code where the docs are silent. The MCP line is from the MCP specification's Tools page.

Where does a person approve in Sume's video agent?

In the Agents chat. Sume's docs say to use the Agents tab when a person should stay in the loop: brief, review, iterate, then save the result as a reusable recipe. The agent picks the models and asks before it spends, and the thread keeps every artifact and approval. A brief can add its own checkpoint, such as asking for a draft before spending credits on the final render.

What happens when nobody is watching?

Over the API, nobody is there to answer. A Format run is one unattended turn: it does not stop to ask a person anything, approvals a chat-authored recipe would request are pre-granted, and the run carries on to the paid step within its spend cap. In current code this is text added to the run's prompt: runs started from the Agents UI keep their approval stops, while API and scheduled runs are told that nobody will answer a question and that approval gates are already granted.

When the agent genuinely cannot go on, the run ends failed, never a half-finished completed. The error code unattended_blocked means it hit a gate it could not pass without a person, such as no avatar matching the brief or a missing input it would have asked about in chat. Fix the input or the brief, then retry with a new Idempotency-Key.

What stands in for the approval?

A spend cap. Agent Completions make it mandatory: generation_spend_cap_usd has no default, because the chat's interactive spend-approval prompt is not available to a backend caller, and the docs call the cap its substitute. Spend caps for unattended AI agents covers how each surface enforces one.

How do I keep a person in the loop outside the chat?

Put the approval between runs or calls, not inside one:

• Review, then continue. A Format run is one agent turn. Show its result to a reviewer, then send previous_run_id on a new run with their notes as the instruction; the agent is replayed what it produced, so it can redo one part and leave the rest alone. The continuation is a new run with its own spend cap, as Regenerate one scene of an AI video shows.
• Approve stills first. Avatar video previews generate the first-frame stills without starting the full talking-video render, so a person can approve composition before a full Avatar Video generation is spent.
• Gate tool calls. On Sume's hosted MCP, an OAuth session sees only read-only tools until the user grants Write at consent, and dry_run=true previews admission and cost without submitting. The required idempotency_key is for deduplication, not human approval. The MCP spec says there SHOULD always be a human in the loop with the ability to deny tool invocations.

Human in the loop or autonomous: which should I use?

Sume's docs split it by stage. Interactive chat is the right surface when a human should stay in the loop, which is where a recipe gets shaped. Once the recipe is fixed and only the inputs change, a Format run is the path, and the docs say to build partner integrations on runs, not on chat threads. A person approves while the work is still being defined; a spend cap and an honest failed status take over once it repeats. For the MCP side of the same question, see Safe automation for AI agents that call paid APIs.