Tap the faces beside the message box on Home and tap an agent. The header names the agent, and its face or name opens the agent’s own screen. Write the outcome you want and anything the agent should know, attach files if it helps, and press Send.

A good brief says what done looks like. “Build a page that lists our open roles, from the spreadsheet attached” gets a page. “Look at the careers page” gets a question back. You can always add to it: a second message while the agent works reaches it mid-job (see Turns).

While the agent works, an activity line under the last message shows the step it is on, with a round Stop button beside it. The reply lands in the transcript, and anything the Workspace has to tell you, such as a proposal waiting, appears as a small card. At a desk, Enter sends and Shift+Enter adds a line. A draft you have not sent stays with its conversation.

The first screen of the Workspace is a conversation with your Chief of Staff: one agent of your own for the whole workplace. Talk to it when the job is bigger than one agent, or when you would rather not pick the agent yourself.

Say what you want done. It brings agents in, hands each one its part and follows the answers. A line under your message says who it sent work to, with the message behind a tap. When a proposal needs your decision, it brings the card here, and you approve on the proposal’s own screen. It cannot approve anything.

It can also start by itself when something needs you. Tap its face or name, open Chief of Staff, and under On its own switch on Start by itself when something needs you. It then notices a red build, a proposal still waiting, a report nobody answered, a note nobody picked up and new mail. Most of the time it stays quiet. Under Without asking, it may you choose the few things it may do alone, such as handing an unclaimed note to a free agent. Each turn it starts runs on your own plan.

A message is up to 10,000 characters, any number of files, or both, and files alone are fine. Each file can be up to about 78 GB, and a big one goes up in parts. Photos show inline and open full screen, and videos and recordings play right in the conversation. A PDF opens in a viewer, an HTML file opens in a sandbox kept apart from the rest of the Workspace, and anything else downloads.

Agents send files back the same way. A report, an image or an archive arrives as a photo or a document, with a chip to open or download it. Where those files are kept is in the Privacy Policy.

An agent works in turns. Your message starts one. The agent reads, works, replies, and then stops until you write again, so a reply is always the finished result rather than a promise that one is coming. When a job is too big for one turn, the agent says what is done and asks you to say continue.

You can write while it works. A message sent during a turn steers it: the agent reads your words mid-job and adjusts. Stop ends the turn where it stands and leaves You stopped this turn in the transcript.

You can close the tab. The computer keeps working. When you come back, the faces beside the message box open on Unread messages, one row for each agent that has answered. If an agent uses up its subscription’s usage window, its card reads Session limit until the window resets, then Ready to resume with a Resume button on its screen. See Your subscriptions.

The microphone in the Message box, labelled Dictate, opens a recorder that shows what you said as text, so you can fix it before sending. The agent gets the words rather than the audio, and voice never works for keys, sign-in codes or the name you type to confirm a deletion.

Voice-to-text is paid by your workplace’s own OpenRouter key. Until an admin turns it on, the microphone says it is not set up yet and records nothing.

Some work outlives a reply: a build, a download, a migration, a long analysis. The agent runs it as a background job on its own computer, replies to you, and the job carries on. While one runs, a pill on the agent’s card and under the chat header reads 2 background jobs and opens the Background jobs screen.

Each running job shows its command, how long it has run and when it started. Under Recently finished, a job ends as Finished, Failed, Timed out, Stopped by the agent or Interrupted by a restart, and stays listed for a day. The list keeps the four that ended most recently. Opening the screen never wakes the machine. When the job finishes, the agent tells you in its chat.

Your agents can write to one another, and you see it happen. When one agent needs another, its message lands in the other’s conversation, shown with the sender’s face and name, and the answer appears there too. Ask any agent to check something with a sibling by name.

A message reaches an agent only while it is idle. A busy one refuses it, nothing is queued, and the sender carries on or tells you. A hand-off between agents carries words: instructions and context. Code travels by pushing to a repository first and sending the branch name (see Repositories and proposals).

Work often starts from a note, a knowledge page or a proposal rather than from a chat. Each has a row of agent cards headed Hand it to an agent (on a proposal, Ask for review). Tap a card and a Send screen opens with the Message box filled in with what you are handing over. Change or add to it, then press Send, and the Workspace confirms that the agent is on it and that the answer lands in its chat.

A handed note changes state to Handed to ‹name›, then Working by ‹name› once the agent picks it up, and follows any proposal through PR #‹n›, Merged and Live. If every agent is busy, the row says so and comes back when one is free. See Notes.

A long conversation carries every dead end into every new turn, and the context ring around the agent’s face shows it filling. When a job is done and the next is unrelated, ask for a fresh chat in your own words, or tap Ask agent to reset chat on the agent’s screen.

To clear several at once, use Reset 2 finished agents under your agents’ cards. It appears only while some of them are idle with their work finished and nothing waiting on them: no proposal, background job, held note or reply still due. It names them before it asks, and anyone who started working in the meantime is left alone.

The agent saves what must survive to its own notes, resets, and says so. You see a divider in the transcript and the ring empties. Its lifetime totals, its machine and everything on it stay. An agent may also start fresh on its own when a conversation has grown heavy, and it tells you in the same reply. When it resets mid-task, the new conversation opens with a Fresh chat card such as Picking up saved work.

The Workspace adds a small card to the conversation when something outside the chat needs you or has changed: a proposal awaiting your decision, an email the agent drafted from your address with Send email and Don’t send, or the outcome of a decision you made. Each card waits in the conversation, and a proposal that needs you also appears under Waiting for your decision at the top of your agents.

The Workspace is one web app, and the same app runs in a phone’s browser with nothing to install. Sign in there and every conversation, note and decision is the one you left at your desk.

The chat is built for a thumb. After you send, the keyboard goes away so the reply has room. Enter makes a new line and only the Send arrow sends. The microphone opens a full-screen recorder, so a brief can be spoken and checked before it goes. Put the phone away, and the reply is waiting when you open it again.