Chat

Chat is the page you will spend most of your time on: one conversation with one of your agents, and a preview pane that opens beside it when there is something to look at. It is a working surface rather than a message box: everything the agent does while it answers, the files it reads, the pages it builds, the permission it asks for, appears in the same thread.

The composer

The box at the bottom is the composer. Type, press Enter, and the reply streams in as it is written. If a send fails before it reaches the queue, your words go back into the box rather than being lost.

A split chip in its control row shows which model the chat is running on. Click the name to open the model picker, and the level beside it to change how hard the model thinks, on models that allow that. Until you set one, the level shown is the one the route gives that model, tagged route; your own choice sticks to the chat and wins until you switch to a different model, which starts it afresh on that model's route level. Beside the chip a ring shows how full the conversation has become; clicking it offers to compact the thread or start a fresh one. Conversations explains when to bother.

Slash commands live here too. Type / and a short list appears: /model, /compact, /context, /usage, /status, /stop and more, with /commands listing the lot.

Drafts are kept per chat, so you can walk away and your unsent text is still waiting tomorrow.

Sending more than words

Files of any kind go in through the paperclip, a drag onto the chat, a paste, or a tile dragged from the preview pane. Thumbnails appear above the input with progress and a cancel, and an upload interrupted by a reload resumes. Attachments are kept in the agent's workspace, so size is never the reason a message fails to send. Paste a long block of text and it becomes a Pasted text card on the message instead of filling the box.

The microphone records a voice message, transcribed on arrival and sent with a player and its transcript. The telephone beside it starts a live voice call with the same agent: that one is still experimental, and says so the first time you use it. Speak back has replies read aloud. All three are covered in Voice.

Send something that looks like a password or a key and Player One stops you first, pointing at the Vault instead.

While it is working

You say: "Actually, make it the Hanley account, not Kerrigan." The reply is already running.

Player One does: a Stop button has appeared and Send has become Add to queue. Your message stacks in a panel above the composer. A short one usually reaches the reply already running rather than waiting its turn, so a correction lands while it still matters, and the row tells you which it will do. The panel can pause the whole queue, resume it or clear it, and a model change you queued can be taken back from its own row. Stop ends the reply and pauses the queue, so what you had queued waits for you rather than going out by surprise.

One line at the foot of the reply says what the agent is doing at that moment: a spinner, a verb and the thing it is working on, such as reading a file, running a command or using a skill. Each new step slides up into place as the last one leaves, and the thing being named is coloured by what it is, so a skill, a tool, a file, a command and a web address are told apart at a glance. Before there is anything to show it says Thinking, rephrasing itself now and then so you can see it is still alive, and where the model streams its reasoning the line carries the latest sentence of it instead. If the connection to your installation drops mid-reply the line says so, and picks the run back up when it returns. If a reply stops making progress, a line takes its place: "This reply looks stuck. Press Stop to end it, then send your message again."

Agents can write their own notes onto that line, so a long piece of work can tell you what it is up to in its own words without cluttering the answer. The line goes when the answer starts to arrive: any prose after that is the reply itself.

Click the line, and What's happening opens: every step so far, filling as they land. It keeps its place while a step is on screen and shows what each one was given and what it returned, with the agent's thinking quoted in full. Once the reply is finished the same panel is What happened, reached from the steps chevron in the footer, so you can go back over a reply of an hour ago and see exactly what was run and in what order. A reply that carries on after its text has landed, waiting on a command, says In progress rather than leaving you guessing.

Behind the composer, a band of small coloured cubes drifts across while an agent works, with more of them as the work gets busier. If you find it distracting, rest the pointer on it for a moment on a desktop and a small pop-up offers a switch to turn it off. To turn it back on, use the Working animation switch on Admin > User (/admin/user), which is also the only place to switch it on a phone. The choice is kept in your browser.

Under each reply is a footer naming the model that produced it and what the turn used, explained in Tokens, cost and caching. Where a chat changed model part way down, a labelled divider marks the spot. Where one answer took the agent several turns, a light rule separates them inside the one reply. When a reply could not run on the model it was meant for, a card says which model served instead and why, and opens out to show the walk it took.

If your installation restarts mid-reply, a card says so and tells you whether the work will be picked up again or whether to send a message to continue.

Cards, approvals and helpers

Not everything in the thread is text. An agent that needs a decision from you, or has results worth laying out, puts a card in the conversation: a choice to tap, a short form, a table, a quiz. A card shows what you answered once you have, and a small cross on its Answered tag lets you change your mind. Cards in the chat has what each kind looks like. A password or key goes into a masked box on a card and straight to the Vault without passing through the agent.

When an agent needs permission to run a command or open something secure, an approval card appears with the decision buttons on it, and deciding it lets that conversation continue. A command card offers Once, Always and No; Stay in control says what each one keeps. A recursive forced delete is confirmed by the agent itself before it runs, and the chat then shows a Confirmed delete line with the command and the reason. A request for more access gets Approve and No. When another installation asks to join a share of yours, a card in the chat it was made from asks you to answer, and a Wake now card waits when a member's agent has sent something while the chat was idle: Sharing covers both.

Work handed to a helper shows as a face in a row beside the composer, one click to read. A face that is working wears a turning glow, and the row folds extra faces into a +N button. Where helpers have helpers of their own, a map icon at the end of the row opens a map of the whole team. A message from another of your agents arrives as its own card with the sender named. The chat's menu has a Team messages, all chats setting, Show, Compact or Hide, for how much of that traffic to draw.

The preview pane

When an agent makes something worth looking at, a document, a page, a chart, it opens in a tab on the right and refreshes as the work changes. Press Files (the folder icon) in the chat's controls to open your files in a tab yourself. On a phone the tabs sit behind a Preview tab beside the chat. See The preview pane.

How it fits

Every chat is with one agent, and the scope switcher above the chat list decides who a new one starts with: see Agents. Each chat sits in a project, which is what your folders, your search and your Trash are scoped by: see Projects.

What a conversation teaches Player One need not stay inside it: your agents save what is worth keeping, so something you explained in March is still known in September. See Memory.

The same agent, with the same memory and the same tools, also answers on WhatsApp, Telegram, Slack and email: see Talk to it.

Where it lives

New chat at the top of the sidebar (/) opens a blank conversation with whichever agent the sidebar is pointed at, its name on the row. The page greets you and asks how the agent can help; click the agent's name to pick another, and a row of example prompts sits below the composer, the middle one dropping into the box when you click it. On a phone the page also offers Photo, Voice message and Call shortcuts above the composer.

Below the navigation your chats are listed with folders first, then pinned chats, then the rest by day. A chat's helpers, the chats its agents started for it, sit around its row when you open it, and the cog at the top of the list chooses which kinds of chat it shows: Conversations has the detail.

The chat's own controls float at the top left of the conversation. Files opens the chat's files in a tab. Share sends files from the chat to another installation: see Sharing. An i button shows the chat's details, with a pencil to rename it. A dots menu holds rename, fork, pin, Hide from sidebar (Unhide once hidden), Move to folder and Move to project. Delete is on the chat's row in the sidebar, and sends the chat to Trash. While a telephone call is live on the chat, a pulsing Live call badge sits ahead of the buttons.

There is no bar across the top of the page. The sidebar icon at the top left, the chat's controls and the status corner at the top right float over the conversation. Click the sidebar icon to hide the sidebar and give the chat the full width, and rest the pointer on the icon to float the sidebar back over the page. The corner holds the bell and the readouts you chose, such as running agents and jobs. On a phone, while jobs are running, a round briefcase button with the count appears in it and opens the list of active jobs.

The browser tab is named for where you are. In a chat it reads the chat's name, then your installation's name; on other pages it reads your installation's name, then the page. With several tabs open you can tell the chats apart.

On a phone the controls fold to suit the screen. The effort level, Speak back and the context ring move into a ⋮ menu beside Send, the model picker becomes a sheet from the bottom, and putting the keyboard away keeps what you typed as a draft.