Human-in-the-loop checkpoints are the safety valve that makes hands-off automation worth trusting: they let an otherwise autonomous Autopilot recipe pause at the moments that matter, wait for a person to sign off, and then resume from exactly where it left off. The point of automation was never to remove people from creative work entirely. It was to remove them from the parts that do not need judgment, and keep them at the parts that do.
Why fully unattended pipelines are a gamble
When you schedule a recipe to run on its own, say, generating a week of social cuts overnight, you trade oversight for speed, that is a good trade for most steps. Boarding shots, scoring outputs, re-rendering a failed clip: none of that needs a human watching in real time. But a few steps in any pipeline carry real consequence. A public claim in an ad. A client deliverable that goes out under your name. A run that quietly burns through your credit budget because a prompt produced more shots than expected.
A pipeline with no pause points treats all of those steps identically. It runs to the end and hands you the result, good or bad. A pipeline with checkpoints knows the difference between a step that is safe to automate and a step that deserves a human glance before it goes any further.
What a human-in-the-loop checkpoint actually does
A checkpoint is a step in a recipe that halts the run and emits a request for sign-off. When the pipeline reaches it, the entire run state freezes: the plan, the assets generated so far, the scores each one earned, and the provenance trace behind them. Nothing downstream fires. The run simply waits.
A reviewer then has three moves:
- Approve, the run resumes from that exact point. Work already completed is not redone; the pipeline picks up where it paused.
- Request changes, the run loops back to the step you specify, so the Director can re-plan or re-generate before reaching the checkpoint again.
- Reject, the run stops cleanly, with the reason recorded.
This is checkpoint governance, part of CoreReflex's Workflow pillar. It is the mechanism that turns a scheduled, hands-off Autopilot recipe into something you can put in front of a brand team or a client without holding your breath.
Where to place checkpoints in an Autopilot recipe
Checkpoints earn their keep at boundaries, the seams between a cheap, reversible step and an expensive or public one. Three placements do most of the work.
After the plan, before you spend credits
In CoreReflex, planning, scoring, and editing are free; only generation costs credits. That makes the highest-leverage checkpoint the one that sits right after the Director boards your shots, roles, camera moves, prompts, and before the PRODUCE stage spends a single credit. A reviewer approves the shot list first, so you never pay to generate a plan that was wrong from the start. If you want the full picture of how the Director arrives at that plan, our walkthrough of the agentic Director loop covers the PLAN → PRODUCE → CRITIQUE → ASSEMBLE sequence in detail.
After the quality gate, before publish
Every shot already passes a quality gate that scores it on prompt match, sharpness, motion coherence, on-screen text legibility, on-brand fit, and claims risk. Those checks are guardrails, not a replacement for a final human read on a finished cut. Drop a checkpoint between ASSEMBLE and the publish-or-hand-off step so a brand or legal reviewer can approve the whole film once, with full context, before it leaves the building.
At a budget threshold
Some recipes are open-ended by design, generate variations until one clears a score, for instance. Pair those with a checkpoint that triggers when projected spend crosses a line you set. The run pauses and asks before it keeps spending. We go deeper on spend controls in our guide to budget governance for creative automation, which pairs naturally with checkpoints in the same recipe.
How to add an approval step to a recipe
Checkpoints are configured as steps in the same JSON workflow recipe that defines the rest of the pipeline, so adding one is a matter of inserting a step at the right boundary.
- Open the recipe in the Autopilot editor. Recipes are programmable JSON definitions, if that structure is new to you, start with our primer on JSON workflow recipes.
- Insert a checkpoint step at the seam you care about, typically after PLAN, after the quality gate, or at a budget threshold.
- Set the gate type so the step exposes approve, request-changes, and reject as outcomes, and define where a request-changes outcome loops back to.
- Choose the reviewer the request routes to, a named person or a role on your team.
- Set a fallback for when no one responds in time: hold the run open, or auto-reject so nothing ships without an explicit yes.
- Save and schedule. The next run honors the pause. The reference for each field lives in the documentation.
Routing checkpoints to the right reviewer
A checkpoint is only useful if the right person sees it. Recipes let you route a sign-off request to a specific reviewer or to a role, so the person who owns brand approval gets the brand checkpoint and the person who owns spend gets the budget one. Because the run state is frozen with the request, the reviewer sees exactly what they are approving, the plan, the assets, the scores, not a vague "approve this run?" prompt with no context.
This is the structural difference between CoreReflex and a generic automation trigger. In a tool built to fire webhooks, a "wait for approval" step is a bolt-on with no understanding of the creative work passing through it, here the checkpoint sits inside a pipeline that already knows what a shot is, what it scored, and where it came from. We compare those two philosophies directly in Autopilot versus Zapier and n8n for creative work, and survey the wider landscape in our roundup of the best AI video automation tools.
Provenance makes every approval defensible
An approval is worth more when you can prove it happened. Each checkpoint outcome is captured alongside the run's provenance: who approved, when, and what exact state they approved. That record sits next to the portable trace every generation already carries, model, prompt, params, score. Months later you can answer "who signed off on this, and what did they actually see?" without reconstructing anything from memory. Our deeper look at provenance and audit trails for AI content explains how that trace is built and why it is replayable. For more patterns like this, the automation category hub collects related reading.
Frequently asked questions
How do I add an approval step to a recipe?
Insert a checkpoint step into the recipe at the boundary you want guarded, most often after the plan and before generation. Set the gate type so it exposes approve, request-changes, and reject, choose the reviewer it routes to, and define a fallback for no response. On the next run, the pipeline pauses there and waits for sign-off instead of running straight through.
Does the pipeline wait for sign-off?
Yes. When a run reaches a human-in-the-loop checkpoint it halts and freezes its state, the plan, the generated assets, the scores, the trace. Nothing downstream executes until a reviewer approves, requests changes, or rejects. On approval the run resumes from that exact point without redoing completed work.
Can checkpoints route to a reviewer?
Yes. A checkpoint can route its sign-off request to a specific person or to a role on your team, and the reviewer sees the frozen run state, what was planned, what was generated, and how it scored, so they are approving with full context rather than a blind yes or no.
What happens if no reviewer responds?
That is what the fallback setting controls. You can hold the run open indefinitely until someone acts, or auto-reject after a window so a stalled approval never lets work ship without an explicit yes. Pick the behavior that matches the stakes of that particular step.
Put a human where it counts
Hands-off does not have to mean hands-tied. With human-in-the-loop checkpoints, you let Autopilot do the relentless work, planning, generating, scoring, fixing, and reserve your attention for the handful of moments that carry real consequence. The pipeline runs on its own right up to the line you drew, then waits for you. Start free with no credit card and build your first recipe with a checkpoint where it matters most.