Tap an agent’s card on Home and its conversation opens. The header has Back, the agent’s face inside its context ring, and its name; face and name both open the agent’s own screen. Under the name is the model line, such as Codex · gpt-6-sol · high, which is a button into the model picker. While the agent has work running in the background, a pill under the header reads 2 background jobs and opens the jobs screen.

The transcript is the whole conversation, with day dividers reading Today, Yesterday or a date. Your messages sit in bubbles; the agent’s replies are plain text on the page. Notices from the Workspace, such as a proposal waiting for you, appear as small cards. A new conversation reads Nothing here yet — say hello. When you have scrolled up, a Newest message arrow takes you back to the bottom.

While the agent works, an activity line shows under the last message: three dots and the step it is on right now, or simply ‹name› is working. Next to it is a round Stop button. Stopping ends the turn where it stands and leaves You stopped this turn in the transcript; if the agent had already finished, you are told so.

The compose bar at the bottom, left to right:

• A stack of faces labelled Your agents, shown only when you have other agents. It opens a sheet listing all of them, grouped by room, so you can switch conversations without going back to Home.
• A paperclip, Attach files.
• The Message box. It grows as you write, and a counter such as 3600 / 4000 appears as you near the limit. The microphone sits inside it.
• The Send arrow.

At a desk, Enter sends and Shift+Enter adds a new line. A draft you have not sent stays with its conversation: leave, come back, and the words are still in the box.

A message is up to 4,000 characters of text, up to five files, or both. Each file can be up to 25 MB. A message may be files alone with no words. A file that is empty, too large, or one too many is refused with a line saying which.

Photos appear inline in the transcript; tap one and it opens full screen, and you can page through every picture in the conversation from there. A PDF opens in a document viewer. An HTML file opens in a sandbox, kept apart from the rest of the Workspace. Any other kind of file downloads.

Agents send files back the same way. A report, an image, an export or an archive arrives in the conversation as a photo or a document, with a chip to open or download it. Files in a chat are stored in your workplace’s own cloud storage once a cloud is connected; the Privacy Policy has the full picture.

An agent works in turns. Your message starts one; the agent reads, works, replies, and then stops until you write again. Nothing happens between turns, which is why an agent will not tell you a result is on its way: it does the work first and replies once with what it found. When a job is genuinely too big for one turn, it 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, rather than queueing them for afterwards. The Stop button ends the turn outright.

Every subscription has usage windows, and an agent can spend one. When that happens its card reads Session limit, and its screen shows Session limit reached with the time the subscription resets. Once it has, the card reads Ready to resume and a Resume button on the agent’s screen sends the word resume, and the agent picks up where it stopped. See Your cloud and subscriptions.

You can close the tab. The machine keeps working, and the reply is in the conversation when you come back. A reply waits where it was written, and Home shows an Unread messages row for each agent that has answered since you last looked.

You can speak a message instead of typing it. The microphone inside the Message box is labelled Dictate. At a desk it records in place with a small live wave, then turns your voice into text in the box, where you can edit it before sending. On a phone it opens a full-screen recorder instead.

The recorder is headed Voice message and reads Listening while it records, with the reminder Speak freely. You can edit the words next. A recording runs up to 90 seconds; press Stop recording sooner if you are done. The words then appear under Review message with the hint Make any changes before you send. Fix anything the transcription got wrong, attach files if you like, and press Send message. Try recording again starts over.

The agent receives the words, not the audio, and it is told that transcriptions garble things: when a request reads oddly it says how it understood you and goes on from there. Voice works in most text fields across the Workspace, but never for keys, sign-in codes or the name you type to confirm a deletion.

A turn ends when the agent replies, and an ordinary process it left running dies with it. For finite work that must outlive the turn, a build, a download, a migration, a render, an analysis, the agent starts a background job. The job runs on the agent’s machine, keeps the machine awake until it ends, and records its output in the agent’s home folder.

When a job finishes, it wakes its agent for a quiet turn: nothing is written to you, and the chat shows no row for it. Two exceptions: a job the agent stopped itself, or one it started asking not to be woken, wakes nobody, and the Workspace Agent is never woken by a job on its own machine. The agent reads the outcome and writes to you only if it promised the result or the result needs you. That is the one thing an agent can honestly promise to come back about, and the reason it can say “I will tell you when the build finishes”. A job that is still running when its machine restarts is marked interrupted, not rerun.

You can see the jobs yourself. The 2 background jobs pill in the chat header and on the agent’s card opens the Background jobs screen. Each running job shows its command, Status Running, Running for, Started and its Job ID. Under Recently finished each job ends as Finished, Failed, Timed out, Stopped by the agent or Interrupted by a restart, with how long it ran and when it ended. A finished job stays listed for a day. Opening this screen never wakes the machine.

Your agents can write to one another. An agent sends a message to a sibling by name, and it is delivered only if that agent is idle. A working agent refuses it: nothing is sent and nothing is queued, and the sender is told to stop rather than retry. Agents read the live roster before sending, so they know who is free.

The message lands in the receiving agent’s own conversation, shown with the sending agent’s face and name, and starts a turn there. The answer appears in that same chat, which is yours to read. The one exception is a request from your Workspace Agent: a room agent answers that back to the Workspace Agent, where you asked.

A hand-off carries words: instructions and context, not files, commits, keys or permission. Code travels by pushing to a repository first and sending the branch name. The receiving agent has no authority beyond what you already asked for.

A note, a knowledge page and a proposal each have a row of agent cards headed Hand it to an agent; on a proposal it is headed Ask for review. Tap a face for the agent’s details; tap the rest of its card to hand over the thing you are looking at. If the room has no agents yet, the row offers Create agent instead, and if every agent is busy it says so and comes back when one is free.

Tapping a card opens a Send screen titled with the agent’s name and the line Workspace conversation. Its Message box is filled in for you with what you are handing over; change it, add to it, or dictate into it, up to 4,000 characters, then press Send. The Workspace confirms with ‹name› is on it — the answer lands in its chat, and that is where to read the reply.

A note you hand over changes state to Handed to ‹name›: given, but not yet started. When the agent picks it up the note reads Working by ‹name›, and if the work ends in a proposal the note follows it through PR #‹n›, Merged and Live. See Notes, knowledge and Drive.

A long conversation carries every dead end and pasted log into every new turn, and the context ring around the agent’s face shows it filling. When a job is done and the next one is unrelated, start fresh. Ask the agent in your own words, reset chat or start a new chat in any language, or tap Ask agent to reset chat on its screen, which sends that message for you and then reads Asked — it will answer in its chat.

The agent does the reset itself. It first writes what must survive into its own notes on its machine, then resets, and says so in one plain sentence. You see a divider line in the transcript, and the fresh conversation waits for your next message. The context ring drops back to empty. The agent’s lifetime Turns and Worked totals are kept, as is its machine and everything on it.

An agent may also start a fresh chat on its own when a conversation has grown heavy and the next task is unrelated; it tells you in the same reply. Nothing resets just because the window is large. When an agent resets mid-task to carry on with work you already asked for, the new conversation opens with a Fresh chat card such as Picking up saved work.

The Workspace adds small cards to a conversation when something outside the chat needs you or has changed. Each is labelled, and most have a door:

• Change proposal: the agent proposed a code change. The badge reads Awaiting approval and Review and decide opens it. Later the card reads Merged with See what changed, Rejected, or Conflict. See Repositories and proposals.
• Knowledge proposal: the same pattern for a knowledge page; an approved page reads Live.
• Site: a site was deleted. The card reads ‹site› stopped serving and notes that its address stays reserved to you.
• CI secrets: the decision on a request to let checks use the room’s keys, Kept private or Allowed. See Keys for checks.
• Turn: You stopped this turn, or Workspace Agent stopped this turn when you asked it to.
• An email card when the agent has drafted a message from your address. It shows the exact From, To, Cc, Subject, files and words, with Send email and Don’t send. Nothing goes out until you press Send; afterwards the card reads Sent to … or Nothing was sent. See Email.

Every card is a record, not an alert. It waits in the conversation, and a proposal that needs you also appears on Home under Waiting for your decision.

The Workspace is one web app at agenticcloudcomputer.com/workspace, and the same app runs in a phone’s browser. There is nothing separate 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. On a touch keyboard, Enter makes a new line; only the Send arrow sends. The microphone opens the full-screen recorder, so a brief can be spoken rather than typed and checked before it goes.

Send the brief, put the phone away, and the reply is waiting in the conversation when you open it again, with Home’s Unread messages row pointing at it.