# Running a brief on autopilot

> In short: On autopilot, CODA plans the work, you approve the plan and its spending cap once, and the run makes the takes, stopping for review where the plan says. It never spends more than the cap. Autopilot runs come with the Studio and Enterprise plans.

Applies to: v0.9 · Last reviewed: 2026-11-02 · Canonical: https://docs.zymcos.com/product/running-a-brief/

On **autopilot**, you describe what you want once and let it run: CODA plans the work, you
approve the plan and its spending limit, and the run makes the takes while you do something
else. It stops to let you review where the plan says so, and it never spends more than the cap
you approved.

There is no switch to flip. Ask CODA to build it for you — "build the whole thing for me,
here's $20", "run this overnight" — or say yes when CODA offers a plan. CODA never starts a run
on its own: the plan card is an offer, and your approval is what starts it.

**Autopilot runs come with the Studio and Enterprise plans.** On another plan, CODA works
through the request with you step by step instead, and says so if you ask it to run something
unattended. One picture or one clip needs no autopilot on any plan: ask for it, even with "just do
the whole thing", and CODA puts it on a card for you to approve.

## The plan card

Ask CODA to build it and write the brief — "run this for me: three moody stills of the
lighthouse at dusk, then a slow push-in on the best one", or simply "a ten-second clip of a
lighthouse at dusk, run it for me". CODA answers with a **plan card**,
and a line above it saying how it read your request ("Reading this as: …"), so you can correct
it before approving anything:

- the brief, as you wrote it;
- the **checkpoints**, in order. A *Generate* checkpoint makes things, one step per take; a
  *Review* checkpoint waits for you;
- what each step is expected to cost, and the **total**;
- the **cap**: the most the run may spend. It is the figure your brief named, or the default.

A step whose model has no recorded price says **unpriced**, and a step no model can take says
**unroutable**, in red; the plan cannot start while any step is either: a run never spends money
it cannot count against its cap. The card says which step it is — ask CODA for a plan on a
priced model. When the cap is below the total, the card says so too: the run can still start,
and stops when it reaches the cap. A step that builds on your review says so.

The card's one button is **Start run**. It creates the run and spends nothing; the run opens in
the **Run** tab, waiting for your approval. Once the run exists, the card's button becomes
**Show the run** — after a reload too — so one plan never makes two runs.

## The Run tab

The right column's **Run** tab lists the production's runs, newest first, and shows the one you
choose. It is always there — a run keeps going while you carry on working with CODA.

For the chosen run you see:

- where it is — *Waiting for your approval*, *Running*, *Waiting for your review*, *Stopping*,
  *Complete* — announced as it changes;
- what it has **committed against its cap**: the money reserved or spent so far, from the
  ledger. A checkpoint with nothing reserved yet says so; it never shows $0.00;
- every checkpoint and its steps, and a failed step's reason in one sentence of ours with its
  code — or, when the step was refused before anything was spent (no model can take it now, its
  prompt is missing or not accepted, or a reference is one the model cannot use), the same code a
  take you asked for yourself would be refused with (see
  [When a take fails](https://docs.zymcos.com/product/when-a-take-fails/#in-an-autopilot-run));
- why it stopped, when it has.

**Approve plan — up to $2.00** is the click that authorises spending: the run may now spend up
to its cap, and nothing more. It appears only while the plan is waiting, and not while runs
are frozen for your workspace. A run's cap is separate from the production's daily limit on
clicked takes: neither counts against the other (see [Spending limits](https://docs.zymcos.com/product/spending-limits/)).

If **Start run** says that autopilot runs come with the Studio and Enterprise plans, your
plan does not include them: nothing was started or spent.

## Stopping a run

**Stop run…** asks before it acts, because a stop is final: a stopped run cannot be resumed —
to carry on, start a new run from CODA. No new step starts after a stop, and the takes still
being made are **cancelled** at once: the run shows *Stopping* for a few seconds, then
*Stopped*.

A cancel stops the waiting; it cannot always stop the bill. If the provider had already started
a take, it may still finish — and then it arrives in your History and is charged. The Run tab
says so, for example *Stopped. 2 takes may still arrive and be charged.* See
[Cancelling a take](https://docs.zymcos.com/product/cancelling-a-take/).

A run also stops on its own when the next step would go over its cap, when a step cannot be
priced, or when a required step fails. The Run tab says which. A run stopped by its cap does not
cancel the takes already in flight: they were inside the cap you approved, and they finish.

## Reviewing a checkpoint

When a run reaches a *Review* checkpoint it waits. Under that checkpoint, the Run tab offers
**Review in the Viewport**: the Viewport's strip then holds only what the checkpoint before
it made, ranked by what you have said about each take —

1. **★ circled**,
2. **👍 liked**,
3. **not rated**,
4. **👎 not liked**.

Rate on the stage as you always do; the ranking follows. The **×** on the banner above the
stage brings every generation back.

Back in the Run tab, the review names **what approving continues on** — the ★ and 👍 takes, by
name — before you click **Approve review**. Approving starts the next checkpoint on those
takes: each step the plan marked as building on your review takes them as its references, or
— when you chose one — as the take it varies. The rest of each step stays as planned. With
nothing marked, the next checkpoint proceeds on the plan's own inputs, and the button says so.

**What the review names is exactly what is approved.** The list counts every take the
checkpoint made, including ones older than the Viewport has loaded, and approving sends those
takes and no others. Approving with nothing marked means *none*: a take marked a moment after
you clicked does not change what the next checkpoint builds on. While the Run tab is still
reading the checkpoint's takes, **Approve review** waits, and says so.

If the choice cannot be used — several takes marked for a step that builds on exactly one, or
more than a step can hold — nothing is approved, and the Run tab says what to change.
