#Chat with an Agent

Chats come in three kinds. Keep them separate:

• Agent Chat is a standalone chat with an Agent or Crew. Use it when you want to ask a question, investigate a topic, shape a plan, or start work without first opening a task.
• Agent chat on a task lives on that task. It is not a task comment and does not require the task to be assigned first.
• Direct messages with a teammate are human-to-human chats — see Direct messages with a teammate. They have no run states, environment preparation, or execution progress; those are Agent concepts.

Each standalone Agent Chat is backed by its own chat session and chat messages. The session belongs to the person who created it. Other organization members do not receive access to that session through the current chat APIs.

#Agent Chat and task Agent chat are separate

Both surfaces can start Agent runs, but they store different records:

• Agent Chat uses standalone chat sessions and chat messages.
• Agent chat on a task is attached to that task. It is not stored as a task comment.

Sending a message in Agent Chat does not create a task comment. Sending a message in the task's Agent chat does not add a message to a standalone Agent Chat session, and it also does not appear in Activity as a comment.

Use Agent Chat for open-ended work. Use a task when the work needs an Assignee, status, review, acceptance criteria, or a durable task record. If the team needs to see a conclusion, copy it into the task description or a comment.

#Start from a task

You can chat with an Agent from task detail without first assigning the task to an Agent or Crew.

The chat opens in a side panel on the task. It does not take one of the Activity / Conversations tabs. The Conversations tab lists every Agent conversation on this task, including ones opened by an @ mention in a comment. The top of the panel has the Call Agent entry and the conversation history for this task:

• the history lists the most recent conversations; further ones are folded into More (N). The block does not appear when the task has no conversations yet;
• Create new conversation in the Call Agent menu at the top of the panel matches the same-named item in the ⋯ More actions menu (icon, label, and behavior) and opens the same Start a new conversation dialog to start an additional chat on this task. One Agent can have several chats on one task without replacing each other, and the roster shows how many chats that Agent or Crew already has here.

The Start a new conversation dialog lets you set up the new chat:

• Select an agent: defaults to the current conversation's Agent, and you can pick another Agent or a Crew in the Space instead. With an Agent selected you can also set this chat's Runtime & model; a Crew has no such row;
• Reuse the current conversation's working directory: with this switch on, the new conversation continues in place in the current conversation's working directory, and both chats share the same files;
• Set a code baseline branch: with this switch on, the new conversation works in an isolated worktree instead of the original folder. Under Baseline branch (optional) you can enter an existing branch; leave it blank to use the current code state. With both switches on, the shared directory runs through an isolated worktree;
• With both switches off, the new conversation uses a fresh working directory. Switches can be unavailable — for example the current conversation has not started, its working directory lives on another computer, or you lack directory permission there — and the reason is shown;
• Enter the first message and click Start conversation.

Assigning a task to an Agent or Crew that already has chats on it — when the assignment starts a run right away — first asks where this run should speak: Run in the existing conversation, Run in a new conversation, or Cancel (cancel means the assignment never happens). Choosing a new conversation makes it the executor's primary chat on the task; later default runs and the Assignee tag follow it. When the executor has no chat on the task yet, there is no prompt: the assignment goes straight through and a fresh chat opens.

Each row in that list also carries the conversation's state — what this conversation needs from you right now. Clicking it opens an actions menu, such as open the conversation, mark as completed, view the run log, retry, stop, or clear the flag; which actions are available depends on the conversation's current state. Mark as completed sets the state to Completed directly and sends no message; marking a task conversation completed also marks the task itself completed. The menu also has a Quick replies submenu that sends a saved reply straight into that conversation. The task header no longer repeats the same state.

Panel titles, button labels, and how many conversations are listed follow the product UI.

The ⋯ More actions menu at the top of the conversation panel offers:

• Create new conversation: the entry above for starting an additional chat on this task;
• Invite members to this conversation: invite another member into the current conversation. Pick a member (an optional note can be attached) and confirm; the invitee receives an invitation card in your direct message with them and joins the conversation once they accept. Any member who can see the conversation can invite, and repeated invitations are allowed;
• Transfer this conversation to another member: hand the conversation to a member. Only the conversation's owner can transfer it — the item is disabled with a hint otherwise. After you pick a member and confirm, the recipient receives the transfer invitation, and the transfer completes only once they accept. The conversation then appears in the recipient's conversation list, the message stream shows that it was transferred from one person to another, and the recipient also receives an Inbox notification, a desktop banner, and an IM notification; if you are also the task's Assignee, you are asked whether to move the task assignee to the recipient as well. Standalone conversations support the same transfer;
• Copy link: copy a direct link to this conversation (task conversations);
• Copy working directory path: copy the current working directory of the lead Agent (for Crew conversations, the Crew leader). The home portion of the path is masked as ~/… (on Windows, %USERPROFILE%\…); while the working directory is still queued or initializing, copying waits;
• Export ACP session: export this conversation as an ACP session file; available only when the conversation's Runtime supports it, otherwise the item explains why it is disabled.
• Turn this conversation into a task: sends the Agent a message (queued while the Agent is busy) asking it to create a task from the conversation's intent and bind this conversation to the new task;
• Bind this conversation to an existing task: opens a search box where you pick the task to bind (standalone conversations).

Opening a chat lands on the newest message by default instead of the top, and earlier content loads on demand as you scroll up.

Each row in the conversation list shows that chat's current state (Working, Queued, Needs your reply, Needs your review, Exception, Completed, Aborted); a stopped conversation reads Aborted, and sending a new message clears it. You can also attach a new state to a reply — any new message clears the conversation's old state. A chat created by a quick action shows the action's name instead of the Agent's.

When the Space has no eligible Agent or Crew yet, the Call Agent menu still appears, explains that none is available, and offers New Agent; when there are no quick actions yet it says No quick actions available instead of showing only the management entry.

When the side panel opens on its own:

• Opening a task no longer auto-opens it. States such as a running Agent, unread replies, or a hand-off waiting on you surface in the task's own conversation list and the sidebar's Task Chats module; open a conversation from the panel's entries, the Activity / Conversations records, or the sidebar module when you need it.
• Assigning a task to an Agent or Crew, and the executor starting a run, still auto-open the panel, controlled by your account settings: Settings → Task conversations → Auto-open conversation after assigning (on by default); auto-open after quick actions has its own switch.
• Links from the Inbox or an @ mention still open the matching task and conversation directly.

While the container is wide enough, the task and the chat open in an even 5:5 split, and the divider between them can be dragged. In a container narrower than 700px the chat opens as a drawer instead; the entries and notices still appear in a single row above the task title, and the mobile drawer has a Back to task entry that returns you to the same task detail without losing context.

The composer also switches the Runtime and the model for this conversation. Where the model publishes one, the model picker can set a thinking depth as well.

When the assigned Agent is working, that Agent's chat is the default. After you change the Assignee, the previous conversation stays in the conversation history of the side panel and can be opened again from there.

New messages in this task chat, along with Agent replies and related handoff notices, are directed to the conversation's owner — normally the person who started it; a transfer moves the notifications to the new owner. When the Assignee started it, that person is the Assignee. They are not broadcast to every subscriber or automatically sent to people who participated earlier. Ordinary comments follow the task's comment and subscription notification rules. @ mentioning an Agent in a task comment also starts a new chat on the task instead of waiting for a reply only in Activity.

An offline Agent is shown as offline. Do not expect a thinking or queued state. When the Agent is offline or starting, or when the composer shows an abnormal status, the notice includes View details so you can open that Computer or Runtime. Task chat, the sidebar, ordinary Agents, the Assistant, and Crews use the same notice. If the Computer is online but the selected Runtime CLI is missing on that Computer, the composer says that computer does not have the CLI, instead of “This Agent was offline”. An archived Agent cannot receive new messages: the composer is disabled. History remains readable. The Assistant conversation still belongs to the assistant; even if the Organization has no other custom Agents, that session does not become an archived-Agent chat.

The composer may also show a background-jobs badge [↑] N, where N is the number of background tasks the current Runtime is running. Only the Computer's owner, working on that Computer, can open the badge to see each task's command and elapsed time; those details stay on the local machine and are never uploaded.

While an Agent prepares to run, the conversation shows live progress such as cloning a repository, creating a worktree, and checking out code. You do not need to confirm those steps in a terminal on the Computer.

Hover an Agent avatar in chat or on a task and use Open Agent settings on the hover card. The built-in Assistant does not have that entry.

#Direct messages with a teammate

Besides Agents, you can DM a teammate directly. Two entry points:

• hover any member's avatar and choose Message on the hover card;
• open Contacts, pick someone under People, and choose Chat.

Rules and details:

• the same two people keep a single conversation: starting again from either entry returns to your existing thread, and an archived DM is revived rather than duplicated. The conversation is created only when the first message is actually sent;
• you cannot DM yourself, and the other person must be a member of your Organization (or a guest sharing a Space with you);
• the title is simply the peer's name — DMs cannot be renamed; the row's context menu only offers Pin to top / Unpin / Hide, and there is no completed state to mark;
• a human DM has no Agent run behind it, so it never shows Working, Queued, or Needs-your-reply states — what it can carry is an unread badge with the number of unread messages;
• messages you send show whether the peer has read them; image-only messages and invitation or transfer cards count toward unread like any other message;
• whether incoming human-DM messages notify you is controlled in Settings → Notifications → Direct messages; desktop notifications are on by default and other channels are off.

#Task Chats in the main sidebar

The main sidebar has a Task Chats module that collects the task conversations you started — tasks assigned to Agents or Crews, quick actions, and conversations opened by an @ mention in a comment — ordered by recent activity. It is the daily surface for "what are my conversations doing", so you do not have to open each task to check on progress. Task conversations started by an automation do not enter this sidebar list; they stay on their task, visible in the task's chat panel, and appear for a person only after that person sends the first message in the conversation. Direct messages with teammates appear in this module too: the row shows the peer's name and avatar, ordered by recent activity like everything else, and carries no run states.

Each row has two lines:

• The first line is the title: a task row leads with the task's status and type glyphs and shows the task title; the Agent or Crew avatar sits at the right edge. Conversations you do not own (for example one transferred to you) show both the owner's avatar and the Agent avatar;
• The second line shows the conversation's current state (Working, Queued, Needs your reply, Needs your review, Mentioned you, Exception, Completed, Aborted), the newest message preview, and the activity time. A hand-off reads "Needs your…" only to the person it waits on; everyone else sees the neutral "Needs reply / Needs review". When someone @ mentions you inside a conversation, the row shows a red Mentioned you, which clears once you open that conversation.

How it behaves:

• Conversations with fresh activity — an Agent starting work, queuing, waiting on you, or an error — rise to the top of the list; a quiet conversation with unread messages shows a small green dot on the avatar instead;
• Hovering a row turns the trailing time into a ⋯ menu button; the menu it opens is identical to the row's context menu — Pin to top / Unpin, Hide, and, depending on the row type and state, entries such as Mark as completed (a human-DM row offers only Pin and Hide);
• Any conversation can be hidden: choose Hide from the row's ⋯ menu or the context menu, and the conversation leaves the list until a new message arrives; hiding also unpins it. Hiding affects your list only — it changes no conversation state, task, or unread, and nothing for anyone else;
• Conversations that stay quiet for a long time (no active state, nothing unread) sink into More (N) at the end of the list, where you can still open them or mark them done; a folded conversation comes back on its own as soon as it has fresh activity;
• Click the module title to collapse the whole module; the collapsed header shows one status glyph per distinct state. The + at the title's end starts a new conversation.

Agent messages and hand-off states in task conversations no longer enter the Inbox — this module is where they surface. The desktop banner and Slack notifications for the conversation starter still go out. Opening a row takes you straight to the task detail with that conversation.

#Execution progress and tool calls

An Agent reply carries the process behind it, such as which files it read, what it searched for, and which commands it ran.

• While the run is active, that process collapses into a one-line summary and only the current step stays expanded. When the turn finishes, the whole process folds into a single summary line that shows how long the turn ran and that you can expand to read every step.
• Reading a file, searching, opening a web page, and running a command each use their own card, with the result shown inside the card instead of as a raw payload.
• To see the raw input or raw result of a tool call, open Raw data from the card. It is not shown in the conversation by default.

When a turn ends abnormally (the run failed or was interrupted), the process output it already produced is kept and stays expanded instead of collapsing, so you can see how far the Agent got; the failure details appear below the process instead of leaving a blank area. A tool that reports an error during an otherwise normal run does not count as an abnormal ending, and that turn still collapses.

#Start and resume a session

Open an Agent's chat view, then send a message or attachment. A new session is created when the first message is sent. The + menu at the top of the home page also offers New Agent chat: pick it below the task types, choose an Agent, and the chat opens — Back returns to the task-type list.

When you return to an Agent with no session explicitly selected, Anakin opens that Agent's most recent active standalone session that contains messages. Choose New session when you want a clean conversation instead.

The session picker lets you:

• open another active session for the same Agent or Crew;
• rename an active session;
• archive an active session;
• open an archived session from history.

Archiving a session shows a light confirmation you can undo. It does not open the chat library to confirm. Archived sessions are read-only. Their messages remain available, but you cannot send another turn in them.

The sidebar layout can prefer Task first or Chat first. The control is under Customize navigation / Layout mode and follows the signed-in account. Exact labels follow the product UI.

A standalone chat has a ⋯ More actions menu in its header, to the left of the new-session button. Its items match task conversations: Create new conversation (start an additional chat in the current context), Copy link, Copy working directory path, Export ACP session, Invite members to this conversation, and Transfer this conversation to another member (the same owner-only rule as task conversations; a transfer completes only after the recipient accepts). Crew and Assistant chats support the same items; the link can be copied once the conversation has started.

Standalone chats also carry a conversation state, just like task conversations: the sidebar chat list and the chat page show the current state, and clicking it opens the same actions menu — mark as completed, retry, stop, and so on. A standalone chat has no task behind it, so marking it completed only closes that conversation and changes no task.

Unavailable Agents can be removed from the chat sidebar. Existing sessions remain available to open.

Task conversations you started also appear in the conversation list. Only the ones you started are listed, and opening a row goes to that row's own task and conversation instead of the first one in the list.

The Assistant conversation shows Local or Remote. Local means the assistant can open files on this computer; Remote means files should be provided as attachments. Ordinary Agent chat does not show this label.

#Conversation working directory

The chat composer has a folder control, Set working directory, so this conversation can use a directory without changing the Agent default.

• Follow the Agent's default: use the default working directory from Agent Run settings. Until you specify a path, the control shows only the folder icon.
• Specify a working directory: choose a folder on the computer the Agent runs on in the Choose working directory dialog. After you set a path, the control shows the directory name. Under the path field you can switch Run in isolated worktrees or Run directly in this folder, or keep following the Agent default.

On the web, or when the Agent runs on another computer, the dialog lists folders level by level through that computer's local service and shows folders only. Use the breadcrumb to go back up, or paste an absolute path to jump straight to a folder. The desktop app on that same computer still uses the system folder picker. Only the computer's owner, or someone the computer is shared with through the Organization or their Space, can browse and set a directory; without that permission the entry point is not shown. Protected locations such as the home folder and system directories can be opened, but cannot be the working directory themselves; on Windows a local drive root (for example E:\) can be used directly.

When the computer is offline, its local service does not respond, or the local service is too old to browse folders, the dialog explains the specific reason (the folder does not exist, is not a folder, the local service cannot read and write it, and so on) and keeps Enter path manually as a fallback. Configure the Agent default in Agent working directory.

#Quick actions

Task detail also has Quick actions in the header toolbar (near star and copy ID; the icon follows the product UI). A quick action runs a chosen Agent or Crew with the current task as context; it is separate from standalone Agent Chat and the task's ordinary Agent chat.

The menu lists the quick actions you can see. The bottom item is Manage quick actions.

Each quick action has:

• Name
• Prompt (optional; if empty, only the task context is sent)
• Agent / Crew
• Visibility: Private (only you can see, trigger, and edit it) or Shared (everyone in this Space can see and trigger it; only the creator can edit it)

Selecting one opens a split view: task on the left, chat on the right. The prompt, the chosen Agent or Crew, and the current task context are sent together. You can keep asking follow-up questions on the right, or pick another quick action.

Each quick-chat trigger creates its own task-level conversation through the current Task's conversation flow. The conversation and its execution record stay on that Task; it does not reuse the assigned Agent's chat, or a chat opened by hand through Create new conversation. Reopen the record from the Task whenever you want to review the context or continue the discussion.

Refreshing restores the last selection without sending it again. Existing conversations can be reopened. An offline Agent cannot send; the composer says so. Deleting a quick action does not delete conversations already created from it.

#Share completed conversation content

Use the share control on a completed response to enter selection mode, then select one or more conversation items to share. When the UI shows a related prompt, image, or attachment with the selected response, it follows the selected content as shown. A response that is still generating, stopped, or incomplete cannot be shared.

Choose a link visibility when you create the share:

• Anyone with the link: anyone who obtains the link can access it.
• Organization members only: only signed-in members of the current Organization can access it.

You can then Copy link, Copy text (the selected conversation as plain text, one block per speaker with Markdown kept as-is and attachments reduced to file names, ready to paste anywhere), or Share image. A share is a creation-time snapshot: later messages, edits, and new content do not change the shared page. Revoke sharing invalidates every active share link for that source.

Review the selected content before sharing. Do not publicly share passwords, tokens, API keys, private keys, personal data, internal logs, or unredacted screenshots. Treat an Anyone with the link URL as externally accessible.

#Send messages and files

The composer accepts text and file attachments. A message can contain only an attachment when no text is needed.

Agent output streams into the session while a run is active. You can stop the active run. If you send another message while the Agent is already working, Anakin keeps the message in the session and queues it for the next turn instead of starting a parallel duplicate run for the same Agent.

Unsent draft text is retained per Agent or Crew conversation when you switch between chats in the current browser.

When an attachment cannot be previewed directly, use Download rather than Open. Anakin shows download feedback, and the desktop app opens the download directory and locates or selects the file after the download completes. This depends on operating-system permissions and desktop-app capabilities; if the file is not located automatically, find it in the system Downloads folder.

Agents can also send files in chat. Attachments that cannot be previewed use the same Download action. When the conversation started from Slack, files the Agent sends also return to that Slack conversation. Older Slack bots may need to be reconnected or created again before they can send files; follow the in-product prompt.

#Quick prompts

The chat composer can use the same personal shortcuts as task comments. Click the prompt text to send it. Hover to show Insert, with the hint Inserts into the composer without sending. Inserting writes the text into the composer and focuses it without sending. Add, edit, reorder, or remove these prompts from the picker; they are stored per person.

#Agent access and availability

You can start a session only with an Agent or Crew you are allowed to use. The target must still be active and available when you send each turn.

If an Agent is archived or is no longer accessible to you, the existing session history remains readable but the composer is disabled. Start a new chat with an accessible Agent to continue.

#Fork a conversation from any message

Any historical message — an AI reply turn or one of your own — can be forked into a new session: the new session starts at that message as the branch point, inherits the full context up to it, and continues in a different direction from there. The original session is untouched, and both sessions can continue independently. Use it to explore a direction suggested by one AI reply, or to try several approaches in parallel from the same context.

Forking works for task conversations and standalone chats alike, on every runtime. A forked session is an independent session; if the original conversation is bound to a task, the fork stays associated with that task.

#Move from chat to tracked work

When a discussion produces a concrete action:

1. Summarize the goal, scope, constraints, and acceptance criteria.
2. Create a task from that summary.
3. Assign the task to the appropriate person, Agent, or Crew.
4. Continue task-specific feedback in the task's Agent chat, or write team-visible conclusions in Activity comments.

The original Agent Chat session remains separate history. Copy any context that the task must retain into its description or comments.

#Common questions

#Can I chat on a task that has no Agent Assignee?

Yes. Open the task and start a chat with an Agent in the Space. Assignment is not a prerequisite.

#Are task Agent chats the same as comments?

No. Comments appear in Activity for the team. Agent replies and related handoff notices in a task chat are directed to that conversation's owner — normally the person who started it; a transfer moves the notifications to the new owner. When the Assignee started it, that person is the Assignee. They are not broadcast to every subscriber or earlier participant.

#Can I DM a teammate directly?

Yes. Hover a member's avatar and choose Message, or open Contacts, pick someone under People, and choose Chat. The same two people always keep a single conversation, and human DMs never show Agent run states such as Working or Queued.

#What happens if the Agent is offline or archived?

An offline Agent is marked unavailable instead of thinking or queued. When the Agent is offline or starting, or when the composer shows an abnormal status, the notice includes View details so you can open that Computer or Runtime. If the Computer is online but the CLI is missing, the composer says that computer does not have the CLI instead of saying the Agent is offline. An archived Agent cannot send or receive new messages; the composer is disabled. Existing history stays readable. Unavailable Agents can be removed from the chat sidebar; existing sessions remain available to open. Archiving the last custom Agent does not turn the Assistant conversation into an archived-Agent session.

#Can other people see my private quick actions?

No. Private quick actions are only visible, triggerable, and editable by the creator. Shared quick actions are visible to members of the Space, but only the creator can edit them.

#If I delete a quick action, do its conversations remain?

Yes. Deleting the quick action removes that saved configuration. Conversations already created from it stay.