Who approves the render: AI copilot or agent for avatar videos?

Let an AI agent submit the avatar preview, and keep a person as the approver of the paid render. Tavus's October 2 piece, AI copilot vs. AI agent, read 2026-10-06, says to choose a copilot when human judgment is the deliverable or the actions are irreversible, and an agent when no user waits on screen and there is a clear completion condition. A finished avatar video fits both: the preview is reversible, and the final render is not.

The article's two lists

Where Sume puts the gates

Sume's hosted MCP lists read tools (avatars_list, avatars_search, avatar-videos_get) apart from paid ones (avatars_create, avatar-videos_create, avatar-video-previews_create, _regenerate and _generate_video). Paid tools need an idempotency_key. The tools-and-gates guide recommends inspecting a tool first with tools_schema and a dry_run=true call, then submitting with an optional max_spend_usd.

The preview adds a second gate that is specific to avatars. It generates first-frame stills before any full render, so the approver looks at the composition, not at a completed video. Changing the script, avatar, scene or aspect ratio needs a new preview. Changing only the quality tier does not.

A workable split

• Agent: draft three script options, create the preview, report the stills.
• Person: pick the script and the still, and confirm the spend.
• Agent: call generate-video on the approved preview, poll, and attach the result.
• Person: watch the finished video before it is published.

Caveat

Spend caps and dry runs are arguments the model writes. They are not a limit that your configuration enforces. If a hard limit matters, put it in the person's approval step.

Review costs less than a re-render. A preview approved before paying for the full video means the person sees the composition once, and the agent never has to guess what was acceptable.