How the agent checks its work

Before saying it's done, the agent critiques the cut, verifies hard constraints, and watches a render of your edit. A gate makes it fix failures.

The agent doesn't hand you an edit and hope. Before it finishes any full edit, it runs a mandatory check sequence — a structured critique, hard verification, and an actual watch-through of the rendered cut — and a quality gate stops it from concluding while any of those checks fail. You get the confidence of a second pair of eyes without doing the QA pass yourself.

You can also invoke every check directly:

Run a review of the edit.

Critique this edit before I export — anything that would embarrass me?

Verify the cut is under 30 seconds and has captions.

Review the edit with a focus on the hook — would it stop the scroll?

Something feels off about the pacing. Watch it and tell me what to fix.

The three checks

critique_edit — the creator-basics review

A read-only review of the current edit against creator fundamentals: is there content on the timeline, is the canvas right for the platform, are captions present, is there a hook, are overlays used correctly, plus any quality and caption risks surfaced by visual analysis, and duration against the target. It can review through different lenses — social (the default), podcast, ugc-ad, product-demo, story, or music.

It also runs a deterministic lint that catches the classic mechanical mistakes:

Overlapping text — hook text sitting on top of captions, or two text blocks colliding
Text out of frame — anything outside the visible canvas
Music-bed coverage — a music track that dies partway through the cut instead of running its full length
Machine-gun cuts — too many sub-second fragments instead of coherent segments

verify — the hard constraints

Quick pass/fail checks on concrete facts: is the duration at most the target, do captions exist, is the canvas portrait, is there media in the project, is the timeline non-empty. No judgment calls — these either pass or they don't. The agent runs verify last after its edits so the outcome is confirmed, not assumed.

review_edit — an AI watches your cut

The most human-like check: the agent renders your current edit and an AI watches the video like an editor would. It comes back with an honest review — concrete problems (caption readability, pacing dead spots, audio issues, a weak hook), the top fixes, and a ship / no-ship verdict. It's read-only; the agent then makes the fixes with normal editing tools and re-reviews until it passes.

You can focus it: "review the edit, focus on caption readability."

The quality gate

The finish sequence is critique_edit → fix what it reports → verify → review_edit → fix and re-run until it passes. The quality gate enforces it: if the agent tries to conclude a run while its latest critique, verify, or review is still failing, the loop refuses to end the turn and bounces the agent back with the failing results — it has to fix the issues and re-run the checks, not re-describe the edit as done.

The gate bounces the agent at most twice per run, so an impossible check can't loop forever. If a check genuinely can't be satisfied (for example, the target duration is shorter than the essential content), the agent stops and tells you exactly what's wrong and why instead of shipping a bad cut.

What you see

Each check appears as a tool row in the chat — the check name, a pass or fail status, and expandable details. After review_edit, the agent summarizes the watch-through in the conversation: the problems found, the fixes it's making, and the verdict. The gate itself is invisible; what you notice is the agent going back to fix something instead of declaring victory early.

When to do it manually

Nothing replaces watching your own cut in the preview before you export — the agent's review is a strong filter, not your taste. Use the checks to catch mechanical problems so your own watch-through is about judgment.

Limits

review_edit skips timelines longer than 120 seconds — for long-form edits, rely on critique_edit and verify, and check sections with the agent ("watch just the first 60 seconds' worth" won't work; trim a copy or review manually).
Rendering for review_edit takes real time — expect it to be the slowest step of a run.
All three checks are read-only: they report problems but never change the timeline. Fixes happen as separate, undoable tool calls.
The gate allows at most 2 bounces per run — after that, the agent reports the remaining failure instead of silently retrying.