Conversations#

A conversation is one piece of work you hand to your agent. Writing, research, organising files, working in your connected apps or on your computer: it all starts in the input box.

Starting one#

• The input box on Home (the row above it lets you pick a knowledge base; "No knowledge base" means none), or New conversation at the top of the sidebar;
• The input box on a knowledge base page: a conversation started there remembers that knowledge base. Your agent looks through its documents, the items filed in it and the watches linked to it before anything else, and documents it writes go there too;
• Email and Telegram: see "Where to reach it" in Your agent.

When a conversation isn't in any knowledge base, documents it writes go to "My knowledge base" and everything else it produces goes to the library.

The input box#

Enter sends, Shift + Enter adds a line.

• +: Upload files, Library (bring a recent item into this message), Recent conversations (reference an earlier conversation as background), Use a skill (pick a skill), Work in a folder on your computer, or Set up an automation (on Home, in a knowledge base and in an existing conversation: write what should happen and when, and your agent sets it up). You can reference up to 5 library items or conversations in one message.
  • Files can be images, PDF, Word (.docx), Markdown and plain text, up to 10 per message. You can also drag them in or paste them (a screenshot, for example). They are saved to your library and your agent reads them before it answers. For videos and podcasts, paste the link.
  • After you choose "Work in a folder on your computer", a tag with that folder appears above the input box; click × to remove it. See Your computers.
• The connections icon: the apps available for this turn, and "Connect more". See Connections.

Start right away, or confirm first#

When you start a new conversation from Home or a knowledge base page, most things start right away. A card titled "Here's how I can help" appears, and waits for you to confirm, in two cases only:

• the result keeps running: a watch or an automation;
• it costs a lot: deep research, or transcribing audio or video.

The card says roughly how many minutes and credits each option takes (or "Free"), and the first option is marked "Recommended". Use the up and down keys to switch, then press Enter or select "Start"; select "Just chat" to skip it, or "Back to editing" to change your message. In an existing conversation the card doesn't appear.

Progress cards#

Setting up a watch, deep research, transcription and imports take a while, so they appear in the conversation as a live progress card. You can leave. When the work is done, your agent picks up in the same conversation (after a transcription, for instance, it goes straight on to the summary you asked for) and sends you a notification.

Open a progress card for the plan, each step, what was produced and what it cost. You can select "Add a requirement" to write one only while a watch is being set up or deep research is still running. Anything still running can be stopped with "Stop" (it asks once more, and finished work stays). "Try again" appears on the page that opens from the progress card when setting up a watch or a research run failed and can be started again from the same request. Transcription and imports have no place to add a request, and Try again does not appear if they fail. When something didn't work, one sentence says why.

What your agent did is one click away#

• Change card: when a turn creates or changes something (a document, a watch, a page saved to the library, a transcribed recording), a card at the end of the answer lists each change. Every line has Open, and Undo while it can still be undone.
• Artifacts panel: the right side of the conversation lists everything this conversation has produced or changed, newest first. On narrow screens it sits behind the Artifacts button in the top bar.
• Split view: on a computer, clicking an artifact opens it beside the conversation. On a phone it opens as a full page, and Back returns you to the conversation.
• Credits used: the bottom of the artifacts panel shows what this conversation has used, split into the conversation itself, the background work it started, and transcription.

When it asks first#

Your agent asks as little as it can. Low-risk things that can be undone it just does: reading, searching, saving items, creating documents, and edits or new things you explicitly asked for. For these it stops and shows a card in the conversation:

• sending something to a new destination outside waper: an email, a message, a place other people can see;
• a single action that would cost more than 100 credits (or more than you have left);
• an edit to one of your documents that was its own idea (it shows you the diff first);
• deleting (unless you asked for it in the conversation and it can be undone), and running commands on your computer.

The card's buttons are "Approve" and "Decline". Things waiting for your decision also appear on Home and in the inbox, and you can answer them on the web (a phone browser works too) and on Telegram. With no response for 7 days, the action is treated as declined. See Inbox and approvals.

The conversation list#

Conversations in the sidebar are ordered by latest activity, and each row shows its state: In progress, Needs you, or New result. A conversation started in a knowledge base carries that knowledge base's name. "View all" in the sidebar opens a page where you can search by title and filter by knowledge base ("All conversations", "Not in a knowledge base"). Conversations from automation runs are not in this list.

The menu at the end of a row (or at the top right of the conversation page) lets you rename it (titles are up to 80 characters, and a title you typed is never replaced automatically), Move to a knowledge base… or Remove from the one it is in (the notice has an Undo that moves it back; documents already written stay where they are, and only later ones go to the new knowledge base). The menu on the conversation page also has Save as skill, and you can delete it (that can't be restored, so it asks first). A message sent while you're offline shows "Will send when you are back online" and goes out once you're connected.

Stopping, and when something didn't work#

If you click Stop, or a turn doesn't work (not enough credits, the model not responding), a line in the conversation says so. Whatever was finished before that still appears on the change card, where you can open or undo it.

Inbox#

The inbox holds what your agent needs from you or wants you to know: things that need you, finished work, reminders, your agent's suggestions and system notices. When something is unread or waiting for you, the inbox icon in the top bar shows a dot; with the sidebar collapsed, the dot sits on the expand button. Each item can be marked done or dismissed, and the Undo in the notice takes it back for a few seconds. Daily briefs never go to the inbox; unread briefs are counted only on Radar. See Inbox and approvals.