Writing a request

The text box is where everything happens. You type what you need and the agent does it.

How to write a good request

State the result you want, not the steps. “Put together a summary of this week’s complaints” works better than “read the complaints and then summarize them”.
Give the context you have in your head: who it’s for, which period, in which tool you want it.
Ask for changes in the same chat. If something isn’t right, write what to change and the agent adjusts it without starting over.
Split big requests. If they’re two different things, send them as two messages.

The text box

Enter sends the message. Shift + Enter adds a line break.
The + button on the left opens the menu to Upload files or photos, Add a new connection or pick a skill.
On the right is the send button and, next to it, the model selector.
The three-dot button, if it appears, holds the chat’s extra options.

While the agent works

You don’t have to wait for it to finish. While it works, the text box says Message the agent while it works…: write a correction or something you forgot (“make it shorter”, “it’s for the sales team”) and send it. The agent reads it at its next step and keeps going with it, without starting over. Stop agent stops it and puts what you hadn’t sent back in the text box.

In a shared workspace, only the person who started the agent can add to it while it works. Anyone else sees “The agent is working on ’s turn. Send yours when it finishes.” and keeps their text in the text box to send it afterwards: that way each person’s messages use their own models and their own usage.

The model selector

Next to the send button you can see and change the model the agent uses in that conversation. Which models are available, and which one a new chat starts on, is decided by your company’s admins (see Which models your organization uses). The No agent option leaves the conversation without an agent, for when you just want to write without it acting.

If your organization already has levels, the selector offers Fast and Thinking, with the model that runs each one underneath; the ⓘ opens Models and levels. Thinking reasons longer and uses more per request. If the chat has images and you pick a level whose model can’t see them, a notice tells you which level to switch to.

Each model says what it’s for. Under them is the Effort: how much the agent reasons before answering. Clicking it opens the levels; the Recommended one is used unless you pick another. More effort usually gives better answers to hard problems, but it takes longer and uses more of your usage; Max uses much more. Your choice is saved for that model. If a model shows no effort, it doesn’t respond to levels.

Suggestions

On the Home screen, under the box, the Get started list shows example tasks. Clicking one copies the request into the box so you can edit it before sending.