Chats

A chat is a conversation that produces work. Everything an agent does happens in a chat, including the runs that start on their own, so you can always read back what happened and why.

There are three kinds of chat:

• Organization chats: a general assistant for your whole organization. Open Chats in the sidebar.
• Agent chats: a chat with one of your agents, which works with that agent's instructions, tools and capabilities. Open the agent from Agents.
• Admin chats: a chat about an agent, to change its settings. See admin chat.

Chats are shared: every member of your organization can see every chat, including the runs of your agents. The Chats list shows them all, with agent chats marked via and the agent's name. Editors and admins can start and continue chats; viewers can read them.

Starting a chat

Write your message in the box at the bottom and press Enter to send it (Shift+Enter starts a new line). Earlier chats are listed next to the conversation: choose New Chat to start another one, or type in Search chats... to filter the list by title. Chats get their titles automatically from their content.

While writing:

• / picks a skill for the AI to use with this message.
• @ in an organization chat mentions one of your agents, to hand it work (see below). In a chat with an agent that has file storage, @ points the agent at one of its files.
• You can attach files from the attach menu, by pasting them, or by dragging them onto the chat. Which files a chat accepts depends on its model. See files.

Organization chats

An organization chat can use every tool of every active integration in your organization, and it does so without asking first: tools never require approval in organization chats. When you want a check before anything is sent, changed or deleted, work with an agent whose tools require approval.

Organization chats always use your organization's memory: they remember facts from earlier chats and can bring them up again. They cannot search the web, run code, create images or keep files; for that, use an agent with those capabilities.

An organization chat can also hand work to your agents. Type @ and pick an agent: it works on the task in a chat of its own, and the result is linked back into your conversation. If the agent stops because one of its tools needs your approval, you get a link to its chat to answer there.

Approving actions

When an agent wants to use a tool that requires approval, it pauses and asks Allow this action?, showing what it intends to do (for example the recipients and text of an email). Choose Approve to let it go ahead or Reject to stop it; after a rejection you can explain what you would rather have.

You can answer whenever suits you, even hours later; the agent then continues where it stopped. If you send a new message instead of answering, the open request is skipped and the agent does not run that tool.

Runs that start on their own, from a trigger or schedule, wait in the same way: open the run's chat to answer.

Answering questions

When a request can be read in more than one way, an agent may stop and ask instead of guessing. The questions appear in the chat as a short form with one tab per question: each has a few options to pick from (one, or several where it says so), an Other choice for an answer in your own words, and room for a note if you want to add something. A check mark on a tab shows it is answered. Answer all of them and press Submit; the agent continues with your answers.

As with approvals, you can answer later, and sending a new message instead skips the questions. Agents ask only in chats you are in: runs started by a trigger or schedule and sub-agents work with what they were given.

Sub-agents

For a larger job an agent can split the work and start sub-agents: copies of itself that each take one part of the task, work on it separately and report back. The agent waits for their reports and then carries on. Several sub-agents can work at the same time, which makes a job like "check these five suppliers" faster and keeps the conversation free of intermediate detail.

Each sub-agent shows up in the chat as one line with its task and its status, for example Working, Needs approval or Stopped, and a check mark once it is done. Open the line to see the prompt the agent wrote for it, everything it did step by step (also while it is still working) and its result.

A sub-agent has the same model, instructions and tools as the agent that started it, and the same rules: when it wants to use a tool that requires approval, the request appears under its line with Approve and Reject, and it waits for your answer. If you send a new message instead, its open requests are rejected and the sub-agent is cancelled.

• Stopping the agent also stops its sub-agents.
• Sub-agents cannot start sub-agents of their own.
• Sub-agents are not listed among your chats; you reach them from their lines in the chat.
• Every kind of chat can use them, including admin chats and the App Builder.
• Their model usage and tool calls are charged like the agent's own.

Models

An organization chat starts on your organization's default model, and an agent chat on the agent's model. You can switch the model of a chat at any time with the model picker next to the message box; the chat keeps the model you last sent a message with.

Long conversations are condensed automatically as they approach the model's limit, so a chat can go on without you managing it.

If a chat's model is no longer offered, a notice above the message box says so and names the model the chat continues with: your organization's default. Pick another model if you prefer.

Seeing what is running

A chat whose agent is working shows a soft green glow on its right edge in the chat list. This covers every kind of run: a message you sent, a trigger or schedule, one agent calling another, or an agent waiting for its sub-agents. The list updates on its own.

In the list, admin chats are marked with a gear and chats started by a trigger with the icon of the connector that started them, for example Gmail.

Focus

The Focus button at the top of a chat hides the navigation and the chat list, so only the chat stays on screen, together with any file preview you open. Click it again to bring them back. Leaving the chat or reloading the page does the same.

Archiving and deleting chats

Each chat in the list has two actions: in the web app and the desktop app point at the chat and use the icons that appear, on the phone swipe the chat to the left. Both need the Editor or Admin role, and neither asks for a confirmation: a notification with Undo appears for a moment instead.

Archive takes the chat out of the list without losing anything. The archive icon in the list's header switches to the archived chats, where Unarchive brings a chat back. Opening an archived chat works as usual, and sending it a new message unarchives it.

Delete removes the chat with all its messages and files. Undo it from the notification while it is shown; after that the deletion is permanent. A chat that is generating a response is stopped first. The chats of sub-agents follow the chat that started them.