# How building with the agent works

> What the agent does after you send a message - questions, the plan, the work, the checks - what it costs, and how you steer, stop and undo it.
> Source: https://playcode.io/docs/start/how-building-works - last reviewed 2026-10-09.

You describe what you want in the chat, and Playcode's agent does the work: it reads your project, asks what it needs to know, plans larger jobs, writes the code, checks the result and shows it in the preview. You steer it with your messages and your answers, and you can stop it or undo its changes.

## What happens after you send a message

The agent first decides what kind of request it is, and acts on that:

| You ask for | What the agent does |
|---|---|
| A change to existing work, like "Make the header dark blue" | Reads the files it needs, makes the change and confirms it |
| A new website whose content, structure or look is still open | Asks questions and agrees the direction with you before it builds |
| A new app with a detailed description of what it must do | Skips the design questions and builds. In a Cloud app, a request with two or more parts gets a plan first (below) |
| A copy of a site, like "Copy this site: https://example.com" | Takes a screenshot and reads the page, then builds from both. It recreates the design, not an exact copy |
| Advice, like "How do I add a booking form?" | Explains first. Say "just do it" when you want the change |

Write in your own words, or dictate with the microphone button. Real details help: give the business name, the services and the contact details you want on the site, and the agent uses them instead of placeholders.

## The plan you can watch

In a Cloud app, a request with two or more parts gets a written plan before any building. The chat shows it as a card, like "Your plan - 4 parts", and the agent stops there. Choose **Build** to start, or **Change the plan** and say what to change.

While the agent builds, the card shows the part in work, like "Part 2 of 4", and "Done" at the end. A part with a page of its own is marked done only after the agent checks that the page answers in the preview. The plan stays in the project's docs, so a later chat picks up where the work stopped.

## How the agent checks its work

In a Cloud app, the agent can run the app's tests and open the app in a real browser to try a flow and look at screenshots. It fixes what it finds and checks again.

In a static site, the preview runs only in your browser, so the agent cannot see it. Tell it what you see, or publish the site so that the agent can open the live address.

Checks find many problems, but they do not prove that everything works. Try the preview yourself before you publish.

## Build quality

Choose how much the agent spends on a task in **Choose build quality**, in the chat's toolbar:

| Level | What the app says |
|---|---|
| **Economy** | "Good for simple edits, fixes, and basic pages." Lowest cost |
| **Standard** | "Beautiful, detailed design. Cleanest results, fewest errors." |
| **Ultra** | "Our most capable model. For the hardest, highest-stakes builds." Highest cost |

On the Free plan only **Economy** is available. Each level has a default model, and you can pick another model for it. See [Models and modes](/docs/build/models-and-modes).

## What a build costs

Every step the agent takes uses credits from the workspace's balance. The cost depends on the build quality, the size of the project and how much of the chat the agent keeps in context. **Context** in the chat's toolbar sets that: **High** is the default, and **Low** or **Medium** keep less history to save cost on simple chats. The app says: "Larger context can cost more, and on some models context above ~200K tokens costs roughly 2x."

Before a big job, the agent tells you what it involves and a starting cost in credits. That number is a floor, not a ceiling. The agent cannot see your balance or what earlier work cost; **Settings**, **Usage** shows that. Generating images, video and audio needs a Pro or Max plan and uses credits too.

See [Plans and credits on one page](/docs/start/plans-and-credits) and [Spend fewer credits](/docs/build/spend-fewer-credits).

## How to steer, stop and undo

- **Answer and redirect.** Send a message at any time to change the direction. Tell the agent what you like and what to change.
- **Stop.** Choose the stop button in the chat (**Stop generation**) when a run goes the wrong way. Stopping early saves the credits the rest of the run would use.
- **Undo, in a static site.** Choose **Undo from here** under one of your earlier messages, change the message if you want, and choose **Undo & resend**. The app says: "All the AI's changes made after this message will be undone, and your project files will be restored to how they were at this point." Then the agent tries your message again.
- **Undo, in a Cloud app.** Restore a checkpoint from **Checkpoint history…** in the project menu. **Undo from here** is not offered there, because a chat cannot undo what already ran on the computers. See [Snapshots and rollback](/docs/publish/snapshots-and-rollback).
- **Continue.** When the agent stops with "I've made good progress but reached the step limit. Click Continue and I'll keep going.", choose **Continue**.

## Background tasks

For heavy work, the agent can start background tasks that work in their own chats while it keeps talking to you. Each task may spend up to its task budget before it pauses. A paused task keeps its work, and only you can let it continue. On Pro or Max, **Settings**, **Billing & plans** shows the budget as "Each background task may spend up to ... cr before pausing"; choose **Change** beside it to set another one.

## What the agent remembers

A new chat is not a new project: the agent reads the files that are already there before it changes them. In a Cloud app, the agent also keeps a written record of the work and its decisions in the project's docs, so the next chat has that context. See [What the agent remembers](/docs/build/memory).

## What stays with you

- The agent cannot change billing or buy credits.
- It cannot read your email, your browser tabs or other sites you are signed in to.
- It asks before actions that are hard to undo or that cost money, like restoring a checkpoint or buying a domain. Nothing happens until you confirm.
- It publishes a Cloud app when you ask. A static site you publish yourself with **Publish**.
- You decide when the result is good enough to publish.