The parts of an AI chat
A chat interface for an assistant has more states than a messaging app. The Chat composer takes the message, attachments, the approval mode, the model and voice. While the assistant works, the Thinking loader shows it is busy and the Thinking state lists what it did before answering. Tool calls get their own card, running, done or failed, with the arguments and result one click away.
The answer itself is an AI answer: the sources it read, inline citations, follow-up questions and the actions it suggests. Under it, Message actions give copy, rate, regenerate and read aloud.
Ask before acting
As soon as an assistant can change data, people need to see and approve what it is about to do. The Approval card states the action, the app it runs in, the risk and the details, with approve or deny and an expiry. The Approvals inbox screen gathers every pending request in one queue, with the trail that led to each.
From a component to a full agent screen
Use the components inside your own chat, or start from a screen: Agent chat is a full-page conversation with reasoning, tool calls, a cited answer and the context window; Agent side panel docks the assistant beside any page so it works on the table with the user.
Running now and Agent monitor, the two status components for background runs, are free with an account. The chat components and agent screens come with Pro.





