Watch it build

Once you approve a step, the agent gets to work and you can follow everything it does. You don't have to watch — the step finishes on its own — but knowing what you're looking at makes it easy to spot when something needs your input.

The chat while the agent works

While a reply is running, the composer shows The agent is working… and a Stop button appears. Above it, the agent's message fills in as it goes:

Short text — what it's about to do, in plain language. The agent is told to avoid technical jargon and file names, and to describe what you'll see on your site.
Activity cards — one small card per action: reading part of the project, writing or editing something, running a check, looking up a guide, searching for photos, taking a screenshot of the preview. Each card shows whether it's still running or done. You can open a card to see the details if you're curious, but you never need to.
The ✅ report — at the end of a step, a short summary of what changed and what comes next.

The preview updates live

On the right, the preview shows your real app running. As the agent changes things, the preview reloads, so you often see the page take shape before the step is even finished. Use Desktop and Mobile above the preview to check both sizes.

The agent also looks at the preview itself: it takes screenshots to check its work, and the screenshot reports any errors on the page so it can fix them before telling you it's done.

Credits used by each reply

Under each reply you'll see how many credits it used (Credits used by this reply). While the agent is still working you'll see a running estimate. How credits are counted is explained in How credits work.

Pauses: the step and credit caps

To keep a single message from running forever (and from spending your whole balance at once), every reply has two limits set by your plan:

a maximum number of steps (actions) per reply,
a maximum number of credits per reply.

On top of that, the agent is asked to keep each reply short — roughly fifteen actions — and to stop at a clean point. When it reaches a limit, it doesn't crash; it stops and writes a notice like this one:

⏸️ Pausing here: I've reached the maximum number of steps for a single message. Everything I did is saved — take a look at the preview, then reply "continue" and I'll pick up where I left off.

To carry on, type:

Example prompt

continue

If a plan is active, you can also click Continue step 2 (with the right number) on the plan.

Note
A paused reply is billed like any other reply — you pay for the work it did, nothing more. If instead you see "Stopping here: you're out of credits", your balance is empty; see How credits work.

Stop the agent yourself

If you see the agent going in the wrong direction, press Stop. The generation stops, what's already done stays in place, and you can explain what you want instead. If you want to undo what it did, use Versions — see Versions: undo anything.

When the agent asks you something

Sometimes the agent stops to ask a question, often with options shown as clickable buttons. Tap one, or type your own answer — the composer says Pick an option above, or type your own answer…. If a plan is active, the plan reminds you to answer the question before continuing.

Tip
Answering quickly and precisely is the fastest way to a good result. "The second option, but in green" is a perfect answer.