Build Agent Overview

The Build Agent is an AI assistant in the Module Builder. Creators use it to build and edit applications and modules, create and use Smart Components, write tests, and answer questions about module configuration. The panel has two tabs: Active Task for the live conversation, and History for all past tasks and plans.

Creators new to the Build Agent can begin with Getting Started with the Build Agent.

Active Task Tab

The Active Task tab is the main interface for sending instructions to the Build Agent and reviewing its responses.

Chat Window

The Chat Window is the scrollable area that displays the full conversation between a Creator and the agent. It includes:

Item	Description
Creator messages	Text the Creator has sent to the agent.
Agent responses	The agent's replies, rendered as formatted text.
Thinking	A spinner that displays while the agent processes a request.
Tool calls	Actions the agent took, shown in collapsible sections. Expand a section to review the agent's actions.
Reasoning	The agent's internal reasoning steps when extended thinking is active.
Checkpoint markers	Saved points in the conversation with a Rollback button. Select Rollback to restore the module to its state at that checkpoint.
Inline plan card	A card that displays in the chat when the agent creates a plan. Displays the plan title, creation date, and status. Select the card to open the plan in the History tab.
Plan approval card	A card that displays after the agent presents a plan for review. Contains two options: Build Plan to approve the plan and begin execution, and Request Changes to ask the agent to revise the plan before building.
Question	A structured question the agent displays when it needs clarification before proceeding. Contains selectable options and an optional text field for freeform input.
Allow/Deny prompts	Confirmation requests for individual agent actions like creating or modifying components. Select Allow to proceed or Deny to cancel.
Stop Build	A button that displays while the agent executes a plan. Select it to stop execution at the current step. After stopping, send a new message to adjust the plan or continue from where execution stopped.
Interrupted build banner	An amber warning that displays when the page reloads during an active build. Select Continue to resume execution from where it stopped.
Cross-application confirmation card	A card that displays when the agent wants to write to a module in another application in the workspace. Shows the target application name, branch, and a direct link to the module. Select Allow to proceed or Deny to cancel. Granting permission covers that application only for the current turn. See Cross-Application Writes.
Lock-refusal card	A card that displays when the agent cannot write to a module because it is not in Editing mode. Contains a one-click action to take the editing lock and retry the request. For a module in another application, the action navigates there first and takes the lock before re-sending the request as a fresh turn.

When no messages have been sent, the chat area displays two starter options: Create an Application and Explain this Workspace. Select either to send a predefined opening message.

Cross-Application Writes

The Build Agent can write to modules in other applications in the same workspace, not only the module the Creator has open. The agent identifies the target, resolves the correct branch for that application, and asks for confirmation before making any change.

Confirmation card

When the agent targets a module in a different application, a cross-application confirmation card displays in the Chat Window before any write executes. The card displays the following:

• The action the agent wants to take and the resource name.
• A disclosure panel with the target application name, a direct link to the module, and the branch name.

Select Allow to let the agent proceed or Deny to cancel. Granting permission covers that application only for the current turn. Permission for one application does not carry over to other applications in the same turn.

Lock-refusal card

When the agent cannot write because the target module is not in Editing mode, a lock-refusal card displays. The card's action depends on where the target module is:

• Same application: The module is in Viewing mode. Select Switch to Editing and retry to switch the open module to Editing and re-send the original request automatically.
• Another application: The module is in another application and is not open for editing. Select Go to [Application Title] on [Branch Name] and take the editing lock to navigate to that application, take the lock, and re-send the request as a fresh turn. The card notes that this action leaves the current view and takes the lock in the Creator's name before re-sending.

If another Creator holds the lock on the target module, no lock-refusal card displays. The agent returns a prose refusal instead.

Turn-level scoping

Once any cross-application write is refused in a turn, all subsequent cross-application writes in that turn are also refused. A refusal can result from a lock, missing permission, or the target being outside the workspace. Start a fresh turn after resolving the underlying issue to retry.

Prompt Field

The Prompt Field is the text input at the bottom of the Active Task tab. Enter instructions and press Enter to send. Press Shift+Enter to add a new line without sending. The field accepts up to 100,000 characters.

Type @ in the prompt field to open the mention popup. The popup lists modules and applications in the current workspace. Type to filter by name, then select an item to add it to the chat context. The selected item appears as a chip in the context strip above the prompt and is included as context in the next message sent.

The context strip above the prompt displays all active context items. Items come from the currently open module, mentions, inline comments on plan steps, and components selected on the canvas. When fewer than three items are active, they display inline. When three or more are active, the strip collapses into a Chat context · N header. Select it to expand. Select the trash icon to clear all dismissible items at once.

The open module joins the context automatically. Select the × on its chip to remove it. This is useful when Creators want the Build Agent to build something new instead of working on what is on screen. It stays out until a different module is opened and returned to, or added back from the module toolbar's Add to Agent Context action, the File Explorer menu, or an @ mention.

When another task in the same module is actively running, the composer displays "Agent is working in [task name]" and input is disabled. Switch to the active task to continue, or wait for it to finish.

The Build Agent does not detect changes made to the module outside the agent during a session. If a component or test is edited manually, include a note in the next message describing what changed.

Agent Configuration

The Agent Configuration button opens a settings popover with two selectors and two toggles.

Mode sets the agent's approach to a request.

The Approvals setting determines whether the Build Agent pauses for review during execution within the current application. Cross-application confirmation cards (see Cross-Application Writes) display separately regardless of this setting.

Auto-test toggle: When set to ON, runs tests after the agent makes changes to verify they work correctly. Disabled when Mode is set to Ask.

Auto-fix toggle: When set to ON, automatically tries to fix issues that tests report. Requires Auto-test to be active. Disabled when Mode is set to Ask.

If a test fails because the Build Agent screenshots the module before it finishes loading, tell the agent to wait longer: "Wait at least 30 seconds before taking the screenshot." The agent respects explicit wait instructions and overrides the automatic 15-second cap.

A risk caption displays below the input area when certain mode and approval combinations are active:

• Build mode: "Agent will build without planning."
• Build + Automatic approvals: "Agent will build without planning and delete without approval."
• Plan & Build + Automatic approvals: "Agent will delete without approval."

Attachments

The Attachments button adds file context to the next message. Attached files appear as chips in the prompt field. Files can also be dragged directly onto the prompt field or pasted from the clipboard.

Files must be 10 MB or smaller total across all attachments in the conversation. Larger files are rejected at upload. The attachment badge in the toolbar turns red at 9 MB used. Additional limits apply by file type:

Supported image formats: PNG, JPEG, GIF, WEBP.

History Tab

The History tab displays all tasks and plans for the current module, grouped by time: Today, This Week, and Older. The Older group is collapsed by default. The tab label displays the total count of tasks.

Search and Filter

Task List

Each task row displays the first message of the conversation as the task title and a timestamp of the last update. An active indicator (a brand-colored dot) marks the currently open task.

Plans belonging to a task display as indented rows beneath it, each with a status badge.

Select the ellipsis button on a task row to open the task actions menu.

When reporting a Build Agent issue to support, download the debug export from the relevant task and attach the zip to the ticket. See Submitting a Support Ticket for what else to include.

Building a Plan

When the agent finishes creating a plan, a plan approval card displays in the Active Task tab. Select Build Plan to approve the plan and begin execution. To skip individual step confirmations, select Approve All when prompted; the agent proceeds through all remaining steps without pausing.

Select Request Changes to revise the plan. Type specific feedback in the chat input or add inline comments directly on the plan steps. The agent incorporates the feedback before presenting a revised plan for approval.

Request Changes is more effective than adding inline comments before selecting Build Plan. With inline comments only, the agent adds them to context and continues with the original plan structure. With Request Changes, the agent rewrites the plan before presenting it for approval. When a plan needs structural changes, use Request Changes instead of adding inline comments and then selecting Build Plan.

The Build Plan button at the top right of the plan detail view performs the same action as selecting Build Plan in the chat approval card.