Agent Chat

Open the AI app and choose New chat. Select a model and begin with a bounded read-only task, such as asking for the available entity types or a summary of the catalog. In a project with the sample store, the suggested prompts above the composer are a good start.

Work in a conversation

1. State the intended task and identify the relevant records or scope.
2. Submit the prompt and inspect tool activity as well as the final answer.
3. For proposed data changes, review the approval before deciding.
4. Check the resulting data in Desk or Explorer when correctness matters.
5. Reopen an earlier conversation from the chat list when you need its context.

The agent operates within the signed-in user’s role and spaces. A prompt does not grant access to hidden data. Model selection affects execution characteristics and cost; use current model and credit information in the app.

If Chat reports that it cannot aggregate, sort, or traverse the data efficiently, the records may already exist and the schema may not yet expose the question. That finding is the start of a model revision — see iterating your schema.

Change the schema from the chat

On projects whose schema is managed in the app, owners and developers can ask the agent to change the data model: “create a schema for a web shop with products, categories and reviews”, “add a rating from 1 to 5 to Review”, “link Order to Customer”. The agent turns the request into schema operations, validates them exactly like a flexkit deploy, and shows a review card listing every entity and attribute it adds, changes or removes. Nothing is deployed until you approve the card; rejecting it tells the agent why.

After deploying a schema on an empty project the agent offers to add a few example records so you can see the app working; it only adds them when you say yes, through the usual mutation review, and tells you they are examples you can delete.

Removing an entity or attribute loses access to its data, so the agent only proposes a removal after you confirm it explicitly, and the card flags it as a breaking change. The deployment then runs like any other and appears on the project’s Deployments page with your name. On code-managed projects the agent explains that the schema is changed with flexkit deploy instead.

Use connected services

Chat can use every plugin connected to the project or to your account, from the first message: “find this week’s emails about the spring catalog”, “which Shopify orders are still unfulfilled?”. If a plugin you need is not connected, the agent shows a connect card in the conversation. Low-risk reads run without a prompt; writes to the external service and anything unclear wait for your approval.

Instructions and memory

Skills provide reusable instructions. Memory is separate from both a skill and the conversation transcript. Avoid treating an answer as proof that every requested operation succeeded; read tool results and errors.

Dictation, where available, also depends on browser microphone permission. Typed input remains the alternative when microphone access is unavailable.

For repeated work with explicit triggers and a run history, create an automation.