> ## Documentation Index
> Fetch the complete documentation index at: https://docs.connie.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Rooms and people

> Create rooms with agents and teammates, have several rooms with one agent, add people with equal room permissions, mention teammates and agents, and know who answers.

A **room** is a conversation with people and agents. Everyone in it reads the same history. Agents answer in the room under their own names, and people see each other's messages with their names. The conversation scrolls vertically without a horizontal page scrollbar; wide tables and code blocks scroll within their own content. Table-only replies use a readable message width, including after you reopen the conversation. Tables stay within the message bubble; scroll sideways inside a wide table to see its remaining columns. Your own messages carry no name. Sender names keep their 13px text size outside the bubble, 4px above it and aligned with its left edge, including agent names. If a reply contains only a question card, the name sits 4px above that card without an empty bubble. Individual messages in rooms and agent DMs do not show a timestamp beneath them. After a pause of five minutes or a change of day, a centered timestamp separates the next message, with 16px of space above and below. Messages have a 16px gap between different senders and a 4px gap between consecutive messages from the same agent or person, including your own consecutive sent messages, artifact cards and account connection requests. Connection-card wrappers add no extra space between these items. It shows **Today**, **Yesterday** or the date, followed by the time. On hover, your message actions appear as unboxed icons to the left of the message. Other people’s messages and agent replies show the same unboxed controls beside their bubble, including when waiting for an answer. For incoming messages, the controls appear as **Reply**, **React**, then the remaining action. If only one other action is available, its icon appears directly; multiple actions use **More message actions** (three filled dots). Your own messages keep the reverse order beside their bubble. **React** opens the full emoji picker on saved messages where you can post. The compact picker has a white **Search** field above eight columns of emoji. Search by name or scroll through the full set; hovering an emoji shows its name below. You can also use the arrow keys and Enter to choose, or Escape to close. Each person has one reaction per message. Choosing another emoji replaces your previous reaction without changing anyone else’s. Click your selected reaction to remove it. The action icons hide when you leave the message, and remain available for keyboard focus or an open picker or menu. Reactions appear in fully rounded badges at the bottom-left corner of your messages and bottom-right corner of incoming messages, including when you expand or collapse a long message. Each badge has even space on either side of the emoji and grows to fit a count when shown. When reactions are present, the normal gap starts below the last row of reaction badges, including when they wrap: 4px before a consecutive message or artifact, and 16px before a different sender or timestamp. Timestamp rows keep their normal 16px gap to the message below. They stay after reload, and you can click your selected reaction to remove it. Hover a reaction badge to see the names of everyone who reacted with that emoji, including when you have read-only access. Open **More message actions** for actions such as copying, editing or deleting a message. In rooms and agent DMs, **Copy message** comes first, followed by any available editing or feedback actions (**Helpful**, **Not helpful**); **Delete message** comes last when you can delete that message. While a room or agent DM loads, its placeholders follow the same centered header, message column and compact message box layout. An empty conversation shows its introductory message centered in the chat area in light gray. Click the centered room or agent name to open **Chat details** beside the conversation: its cover, avatar, name, members, **Tasks**, **Artifacts** and **Context**. Click the name again to close it. Open these details from the centered agent or room name; the agent’s sidebar menu has no **Details** option. **Tasks**, **Artifacts** and **Context** have no **+** buttons in the panel. Hover or focus a room in the sidebar to reveal its three-dot button. Click that button for room actions (hovering over it doesn't open them); with the keyboard, Tab to it and press **Enter** or **Space**. Right-clicking the row does not open the menu. Each option has an icon before its label. The conversation header overlays the scrolling messages with a subtle white fade, from 80% opacity at the top to transparent at the bottom, behind its title and controls. Messages show through the transparent part as you scroll. In a private 1:1 agent DM, the agent’s name and avatar appear in the header and are not repeated beside its message bubbles; group rooms keep sender names and avatars.

Received messages and their artifact cards use at most 80% of the message column’s width in both agent DMs and group rooms. Short bubbles fit their content. In group rooms, the sender’s profile avatar aligns with the bottom of the last visible bubble or card in their consecutive messages, including questions, approvals, app connections and artifacts. It stays aligned when content expands or the window narrows. Reaction badges below the message do not move the avatar down. There is no extra blank space below an app connection card.

The sidebar’s scrolling **Agents** list uses a slim white scrollbar and fades softly at the bottom and, once scrolled, at the top. Search, navigation and the **Agents** heading stay clear, and you can scroll the last row fully above the fade.

Click the account bar at the bottom of the expanded sidebar to open its menu. The menu matches the full bar’s width and aligns with both edges, including when you resize the sidebar. The notification bell remains a separate control.

Room addresses use `/<workspace>/room/<conversation>`. Your private agent DM uses `/<workspace>/agent/<conversation>`. A room keeps its address when people or agents join or leave. Older `/rooms/` links continue to work and open the same conversation if you can read it. Adding people from a DM still creates a separate, empty room; it never shares the DM history.

The reaction picker fits within the browser window and scrolls when the available height is short.

## Create a room

To keep separate jobs organized, use a room for each workstream, such as finding leads, writing outreach or weekly reporting. The same agent can work with you in several rooms. Ask your agent how to separate tasks and it can suggest rooms; asking for advice creates nothing. Ask it to create a room when you are ready. A separate agent is useful for a distinct ongoing role with its own purpose or connection exclusions. Agents use their owner's connections minus any exclusions; creating an agent or room does not connect accounts. A new room starts its own conversation history; it does not copy the old conversation or move existing work. Saved Lists, Records, Pages and Workflows stay saved with their existing access rules.

When an agent's saved reply contains an email, Slack or LinkedIn draft card, **Copy message** copies readable addressing fields and message text with paragraph breaks, rather than the internal draft JSON. Supplied From, subject, CC and BCC fields are included. Rich-text paste keeps safe formatting for an HTML email. Copying the reply does not send the draft.

**1. Open New room**

In the sidebar, click **+** next to **Agents** (**New agent or room**) and choose **New room**. On a narrow screen, click **Open navigation** first, then **+** in the icon rail and **New room**; the menu stays above the dimmed page. Or open an agent’s three-dot menu in the sidebar and choose **Rooms** → **New room with \[agent]**, which creates the room with that agent right away, named after the agent and you (for example "Nova and Yann") so it doesn't read like the agent's own row.

**2. Name it and choose who's in it**

Give it a **Name** (for example "Q4 planning"), pick one or more **Agents** (**Search agents** finds them) and, if you like, **People** from your workspace. Each opens a dropdown list; click outside it or press Escape to close it, and the field then lists who you picked, such as "Assistant, Trading Agent". **Agents** lists only the agents you can use: an agent that is no longer shared with you isn't offered, here or in **Add agents**, even while your DM with it stays in the sidebar. A room you don't name is named after who is in it. The dialog says "All members, including future additions, see the full history."

**3. Create it**

Click **Create room**. The picker stays disabled while the room is created, then closes before the room opens. If creation fails, your selections stay in the picker so you can try again. The room appears in the sidebar with your picture and its agents' faces. If it has agents, one of them starts the conversation with a short message, such as what it can help with here or a question about where to start. A new conversation’s opening greeting appears at the top of the message column and writes out once on its first view. Reopening or reloading shows the saved greeting in full. Reduced-motion preferences skip the animation.

You can have **several rooms with the same agent**, for example one per project. Each room keeps its own history; the sidebar lists each one separately, and a message is always sent to the room you're in.

A room name must contain at least one visible letter, number or symbol. If you try to rename a room using only invisible characters, Connie says "Enter a name with at least one letter, number or symbol." and keeps your draft open so you can correct it.

Your avatar in the room list and room title uses your profile photo. Without a photo, or if it cannot load, it shows your initial on a neutral gray background. Changing the workspace picture or color does not change these personal avatars.

## Audio calls

Click the room or agent name to open **Chat details**, then click **Call** and allow your browser to use your microphone. Hovering an available **Call** button shows “Starts a call with everyone in this room.” and reminds you that text transcripts are saved in **Context** while no audio recording is saved. If calls are unavailable, the disabled button explains that on hover or keyboard focus. Calls support up to 12 people and four active agents. Each agent has its own voice and appears in the call alongside the people who joined. An agent's owner picks that voice in [the agent's profile](https://docs.connie.ai/agents/agents-and-permissions.md#voice).

A shared room with a call shows **Call ongoing · Open to join** in the sidebar. Open it and click **Join call** in its header or details. Only room participants who can post can join; a link or read-only access does not let someone listen. People join explicitly and can join after a call has started.

The call controls stay open when you close details or visit another page in the same workspace. Use **Mute** and **Unmute** for your microphone, **Captions** for live text, **Leave call** to leave, and **End for everyone** to close the call and disconnect everyone in it. Anyone in the call who can post may end it. **Hide captions** closes the text. You can take part in one call at a time. Joining from another tab replaces your previous connection. Changing workspace, signing out or losing access disconnects you.

Agents listen for invitations to speak, including greetings, calling their name, and questions addressed to them or to everyone. When only one agent is in the call, you can also address it as “assistant” or “agent”. In a call with just you and one agent, you can speak naturally without repeating its name. Recognized speech can interrupt an agent's answer; background-audio filtering reduces interruptions from noise. Agents can also ask each other questions and take turns responding. Their automatic exchanges stay brief; ask a follow-up or ask them to continue when you want more. Paused, retired or removed agents do not speak. Calls currently recognize English. Human-only stretches of a call are also transcribed, so agents invited back can catch up on what was said while they were away.

Agents leave the audio call when asked, when they finish an explicitly bounded role, or when their part of the discussion is clearly over. A short pause, an ordinary answer or a simple “thanks” does not make them leave. Use **Ask to leave the call** beside an agent's name to dismiss it, or **Invite back** to bring a departed agent back. Leaving the call does not remove the agent from the room or stop its tasks. **Invite** brings another available room agent into the call, up to four at once. You can also ask an agent to invite a named room agent into the call or invite it back. Add an agent to the room first if it is not available there. Asking one agent to ask another a question starts an exchange between them; the second agent replies to the question it hears. A request to explain the conversation uses the discussion and decisions, rather than a list of the agents’ tasks.

If you ask an agent to do workspace work during a call, it enters the normal task and approval flow without posting a duplicate of your speech as a chat message. The task retains who asked and their exact request. The agent can acknowledge the request while the work starts, continue talking, and speak a concise result when its published reply is ready. Check **Tasks** and the conversation for progress, questions and approvals. You can explicitly ask the agent to approve its current tool calls: plain approval allows those calls once; say **Allow in this session** or **Always allow** to request that wider permission. The app checks the same access and permission rules as the card, and the agent confirms only a recorded decision. Saying you already approved, quoting an approval, or asking another person to approve does not approve anything. Leaving the call does not stop the work. Tasks use the saved call transcript, including relevant later clarifications, so references such as “that deadline” stay available after an approval pause. An interrupted work announcement yields to your next question rather than restarting the old summary. When you ask for collected data in a Records List, the agent should save the actual items and reuse available results rather than collect them again. Voice discussion uses the room's visible context and does not give listeners access to an agent owner's private information.

A finished task from a call publishes its reply without automatically starting a new exchange between agents in the chat. Ask for another contribution in the call or send a message in the room.

Connie does not save an audio recording. During an active call, recent conversation and a summary keep agents aware of earlier facts, including after a coordinator reconnects. This internal working context is erased when the call ends; the saved transcript remains with the room. Spoken work requests remain attached to their tasks, without duplicating human speech in the conversation; calls now also save a speaker-attributed text transcript independently of work requests. When a call ends, it is saved as a Page in the room: an AI summary at the top (an overview, key points and any agreed next steps), then the transcript with each speaker's name and the time into the call. Open **Chat details → Context** and select **Call · \[topic]** to read it beside the conversation, including after everyone leaves, or ask an agent in the room to pull it up. During a call, for a few seconds after it, and for calls from before this change, select **Call \[date and time]** for the text transcript; longer transcripts have **Load more transcript**. Interrupted replies are marked and their unfinished words are omitted. Only people who can currently read the room can read its transcripts. Transcripts are available for calls made after this feature was added; earlier context files contain only the text saved with a work request. When the last person leaves, the call ends and the room's button returns to **Call**; a call left open by a closed or crashed tab closes shortly afterwards. Calls end after one hour. When the call service is unavailable, **Call** is disabled or a connection error appears in the call controls.

### Compare call voices

The experimental **Compare call voices** page is available at `/{workspaceId}/call-demo` in workspaces with agents and rooms. Choose **1–4** agents you can use, then **Create test room** for **GPT Live 1 · experimental**, **Realtime 2.1 · experimental**, or **ElevenLabs + Ultrafast · current**. Each option creates a private room with those same agents. Your other rooms keep their current voice engine.

In the test room, open **Chat details** and choose **Call**. Add teammates through the normal room controls to compare a group conversation. All options share the normal tools, tasks, approval cards, permissions and saved call transcripts. End one call before trying another. You can also **Open this room full screen** and revisit it from the sidebar. Normal tool permissions and usage apply.

GPT Live uses continuous listening and speaking; Realtime 2.1 uses semantic turn detection and low reasoning effort. The native options use OpenAI voices and keep per-person transcription for speaker attribution and authorized actions. They are experiments, not the ChatGPT app. The Live demo uses the available `gpt-live-1` API model; Live 1.1 is not available to our API account yet.

## Add people

**In a room**, open the room’s sidebar three-dot menu and choose **Members**. Under **People**, type in **Add people by name or email** and pick teammates from your workspace. Everyone added has the same room controls: they can post, approve, add or remove people and agents, rename or delete the room, and change its access. Workspace guests can only read. There are no room roles to choose.

Several people can edit one of the room's pages at the same time: if the page changed somewhere else while you were typing, Connie keeps your text and asks whether to load the latest version or keep yours ([Find your pages](https://docs.connie.ai/chat/sharing-and-artifacts.md#find-your-pages)).

The dialog says "People you add can read this room’s whole history, including everything posted before they joined." Click **Invite**. Only people who are already members of your workspace can be added; invite anyone else to the workspace first ("Only members of this workspace can be added. Invite them to the workspace first."). Workspace guests can only read.

Each person you add receives a bell notification naming who added them and the room. **Open** takes them directly to that room. Adding someone who is already in the room again does not send another notification; removing and adding them again does.

**In an agent's DM**, open the agent’s three-dot menu in the sidebar and choose **Add people** ("Starts a new room with \[agent]; this conversation stays private"). Your DM stays private: this starts a new room instead. **Start a room with \[agent]** says "This starts a new room with you, \[agent] and the people you add. Nothing from this conversation is shared; it stays private to you." Pick the people, give the room a **Room name** if you like (optional), and click **Create room**. The new room starts its own history; each active agent introduces itself in turn, without sharing your DM history.

If you ask the agent in its DM to add or invite someone "here," it explains that the DM stays private and offers to create a new room. It does not send a workspace invitation first.

To work with the same agent in a fresh conversation, create a new room and add that agent. Its purpose, tools, connections and authorized memories stay with the agent; the new room starts its own message history. Give it the relevant List or artifact and explain what remains to do. Earlier completed work and your private DM stay where they are. A new room does not cancel work already running elsewhere.

In both dialogs, **How rooms work** at the end of the note opens this section.

Every non-guest participant has the same room controls. They can rename, duplicate and delete it, add people, or remove people and agents in **Members**. This also applies to people who previously had **Post and approve** or **Only read** access. Managing a room does not grant management of its agents or access to anyone’s personal connections.

## Share a room

Open the room's sidebar three-dot menu, choose **Members** and copy the link there. Connie confirms with "Link copied. People in this room can open it." A link never adds anyone to the room. Add teammates under **People** in the same dialog.

Rooms belong to one workspace. Sharing a link does not grant access across workspaces: the person must join that workspace first. After login, someone outside the workspace sees the existing workspace **Access restricted** screen and its **Request permission** action, before Connie tries to open the conversation.

If conversation setup finds that you cannot read the room, or that the link is missing or deleted, it shows **This conversation is unavailable or you don’t have access.** It stops automatic retries, including when you refocus the window, and removes any previously displayed result. The message does not reveal whether a private room exists, its name, or its participants. Temporary connection and server failures still retry. An expired session asks you to log in; a workspace where rooms are disabled shows a separate unavailable message. Opening another conversation or switching accounts starts a fresh access check.

The conversation's **Access restricted** screen, which can also appear when access changes while a room is open, offers **Go back** and **Request access**. **Request access** sends a request to a current non-guest participant. The request says **Approve: Join room**; for a private agent DM, it says **Approve: Start a new room**. After **Grant**, your notification opens the room; that screen also checks again when you return to it. It does not reveal the conversation's name or history, including when a link no longer exists.

A direct message stays private even when its agent is shared with you. If its owner grants your request, Connie starts a new room with you and the agent, and your notification links to that new room. The original DM and its history stay private. If you only need to talk to a shared agent yourself, open its row in the sidebar to enter your own DM.

The header has no **Share** button or workspace-access selector. Existing access stays in effect: a room already open to the workspace remains readable by non-guest workspace members; only its participants can post. An agent's DM stays private.

## Members

Open the room’s sidebar three-dot menu and choose **Members**, to see who is in it: **People**, then **Agents**. The avatars and room name stay together in the same control on a phone. Click **Done** to close it.

Every non-guest participant can also:

* **Remove a person**: click **Remove** next to their name and confirm. **Remove \[name] from this room?** says "\[name] will no longer be able to open this room or read its messages. Someone in the room can add them back." If they have the room open, even in another tab, it closes for them, with any answer an agent was writing there, and shows **Access restricted** with **Request access** and **Go back** The same happens in your other tabs when you leave a room.
* **Remove an agent**: click **Remove** next to it and confirm. **Remove \[agent] from this room?** says "\[agent] stops its work in this room and no longer answers here. Its work elsewhere, its DMs and what it already posted here stay." Its routines that report into this room are paused ("\[agent] is no longer in the room it reports to."), and its event routines here are switched off. Its tasks for this room stop and leave the **Tasks** list, including any it had not started yet, and under each request it was working on the room says "\[agent] was removed from this room, so it stopped working on this." Work another agent in the room had handed it for a request here stops too, with anything it had handed on, and the agent that handed it the work is told that work stopped, so it doesn't post that work as done. Where it had handed the work on in a conversation of the person who asked, it's also told why: the agent was removed from this room.
* **Add an agent**: open **Members** from the room's sidebar three-dot menu (or click **Add agents** next to **Agents**), type in **Add agents by name** under **Agents**, pick them and click **Add**. Only agents you can use and that aren't in the room yet are listed. The dialog says "Agents you add can read this room’s whole history and answer here when they are asked. Adding one starts no work." A room holds up to 8 agents. Everyone in the room then sees a centered **Added agent** (or **Added agents**) line in the conversation, with each new agent's avatar and its name in the agent's color, for example "Added agents Scout and Writer". The line stays when you reopen the room, and the room's agents can read it too. Each newly added active agent also posts a short introduction in its own words, once for that membership. Adding only people does not repeat introductions or start agent work.

Workspace guests and people reading a workspace-open room without joining see the list without these controls. You can't remove yourself; use **Leave room** instead.

## Talk in a room

Agents receive the saved emoji reactions and the names of the people who reacted when they answer in a room or agent DM. When you use **Reply**, the agent also receives the referenced message and its author, including when that message is outside the recent conversation history. This context stays within the same conversation. Deleted or removed messages are unavailable; an edited quote shows its current text and is marked as edited. Reactions and quoted messages do not approve tools or grant permissions.

When an agent answers in a room or agent DM, it reads the conversation's most recent messages, up to about 240,000 characters of text (always at least the last 60 messages). Its short first replies and **Which task is this for?** questions are not part of what it reads. Because each AI step reads this history, a long conversation uses more credits per message than a short one.

A conversation longer than that, or with more than 400 messages, is kept in two parts. Connie writes a record of the older part: what was decided and by whom, texts and drafts in their exact wording, facts and numbers, what was made and where it is, what the agents' tools found or changed, what is still open, and a short timeline. The agent reads that record, then every message after it word for word. The record is written in the background after a message, so nobody waits for it, and it is updated as the conversation grows; each update leaves at least the most recent 30 messages word for word. It does not depend on the agent's AI model, so it stays when you change the model. Editing or deleting a message the record covers, going back to an earlier message in an agent DM, or a change to who can read the room makes Connie write the record again from the messages as they are now. Until the first record exists, the agent is told how many earlier messages it isn't shown.

Nothing is removed from the conversation: you still see every message. The agent can also search the whole conversation for something said earlier and read any message in full, so you can refer to "the copy we agreed last week" without pasting it again.

* **Mention** a teammate or agent: type **@** in the message box, or click **+** at the bottom left and choose **Mention** — **A teammate or agent**. Search the workspace's teammates, agents you can use. The menu opens immediately and shows **Loading mentions…** if the list is still being fetched. Scroll to see the full permitted list, or use the arrow keys to move through it. Press **Escape**, click **Close mentions** (the ×), or click outside the menu and message box to close it without changing your draft. Continuing the same mention keeps it closed; type a new **@** or use **+** → **Mention** to open it again. Keep typing to narrow the list, spaces included: **@Office He** still finds **Office Helper**. Press **Enter** or **Tab** to pick the highlighted name; once what you typed matches no one, the list closes and **Enter** sends your message. In rooms, **everyone** appears first, followed by people and agents already here. Others say **Not in this room**, or **Not in this conversation** in an agent DM. Paused agents also show **Paused**, including when they are outside this room. Private agents you cannot use are not offered. Rooms cannot be added as @ references in the conversation composer. A mention identifies your target; it does not add anyone, send a message elsewhere or share a room's history.
* **Mention everyone:** in a room, select **everyone** — **Everyone in this room**, or type **@everyone** directly. It addresses the current room's participants, and every agent present that can answer receives the message once, even if you also mention one agent or reply to its message. It does not reach anyone outside the room or invite them. This tag is available in rooms only, not in one-to-one agent conversations.
* **Mention or ask an agent outside the conversation:** you and agents can mention another agent in any conversation, including your one-to-one agent conversations. Ask an agent to mention another one, for example "Can you mention Purple?", and it writes **@Purple** in its reply. A mention shows the agent's avatar and name and sends nothing: it doesn't add that agent or share this conversation. To get an answer from another agent, ask for it, for example "Ask @Purple what our refund window is". From a one-to-one agent conversation, the agent sends just that question to the other agent in your own private chat with it, waits for the reply and tells you what it said, or that it couldn't ask (for example because that agent isn't shared with you). Nobody is invited and nothing else from your conversation is shared. In a room, an agent that asks another agent for something and waits for the answer replies to you once: the other agent's own message with the same answer stays in the room for everyone, but it doesn't make the first agent reply again. Writing **@Name** in its own reply doesn't ask that agent anything, so a reply that only mentions an agent and says it asked is checked before it is posted, so the agent can be sent back to actually ask.
* **Reach someone outside the conversation:** if you only speak to an absent person or agent, even with just "@Chief hey", an agent here says they don't receive messages here, and offers to ask that agent for you or, for a teammate, to create a room with them. You can also explicitly ask to create a room and send a message, or post to an existing room you can use. You can select a mention or just type a name; a mention supplies the exact identity, while a typed name may need a lookup or clarification. Adding a person to an existing room gives them its earlier history; creating a new room keeps that history separate. Permissions and approvals still apply. Selected references supply names and IDs, not profiles or another room's messages. If access changes before the agent starts, it does not use the unavailable reference.
* **Reply** to a message: hover it and click **Reply**. The controls stay visible as you move across the gap beside the bubble, including for long replies, and while their menu is open. Your message quotes it, and a reply to an agent's message goes to that agent.
* An agent's answer in a room or DM automatically quotes the message it is answering and shows who wrote it. The quote stays with the answer while it is being written, when it waits for approval or a connection, and after you reload. If the original message is outside the loaded history, it says "Reply to an earlier message" until you load it. Updates from background responsibilities have no message to quote.
* **Edit** a message: hover one of your own messages, click the edit button, change the text and click **Save**. Editing only changes the text: it never undoes what an agent already did with it and sends nothing again. The edit box explains: "Edits this message; replies stay, with no new response." Replies to the earlier wording keep their content, but their quote is marked "(edited)" and says "This reply answered an earlier version." If an agent publishes after the edit and its input version is uncertain, it says "This reply may refer to an earlier version." These quotes do not substitute the corrected wording for what the reply answered. To ask for new work in a shared room, send a new message. You can edit only messages you sent: nobody can edit someone else's message, other room participants and workspace owners and admins included. An edited message shows "(edited)". A message someone else edited before this rule keeps "Edited by \[name]" under the text.
* **Edit and resend** in your DM with an agent: editing one of your messages there goes back to it, like editing a prompt in ChatGPT. Hover the message, click **Edit and resend**, change the text and click **Send**. The box says "Sending removes everything after this message, and the agent answers again. Unfinished work from those messages stops; what was already done in your connected apps, like an email sent, stays." The message and everything after it leave the conversation, your new words are sent in its place (marked "(edited)") and the agent answers them as a new request. Work the removed messages had started stops: queued work, work waiting for an approval, a connection or a reply, and follow-ups they set up. A step already under way, such as an email being sent, may still finish. When going back stopped work, a line under your new message says which, for example "Going back stopped work from the removed messages: “Email Bolt” (asked to stop). What it already did in your connected apps, like an email sent, stays." Nothing is undone: an email already sent stays sent and a record already changed stays changed, and the agent is told what was already done, so it doesn't do it again unless you ask. If you sent another message in the meantime (in another tab, for example), you're asked to reload first: "You sent a message after this one. Reload to see it, then edit again." In a room with other people, editing changes only the text, as above.
* **Delete** a message: right-click the message and choose **Delete message**. With the keyboard, focus the message and use **Shift+F10** or the **Menu** key. In wider conversations, you can also click, tap or activate **More message actions** and choose **Delete message**. Delete never appears as a direct hover icon. In rooms and agent DMs where you can post, you can delete your own messages and agent replies. All non-guest room participants can also delete another person’s messages. Workspace owners and admins can delete another person’s message in any room or agent DM where they can post; this does not grant access to private conversations. Saved agent replies containing only cards can be deleted too. When a reply has no text to copy, its menu still offers **Delete message**. Rooms and agent DMs show no failed replies: when the AI provider has a brief problem the agent tries again by itself, and when it still can't finish it says so in an ordinary message with its usual controls, without an error card or **Try again**; to ask again, send your message again.  **Delete this message?** says "Its text is removed for everyone in this room. Deleting it doesn’t undo anything it already started, and the record of that stays." Click **Delete message**; the message then reads "Message deleted" for everyone. Deleted messages have no **Copy message**, reply, edit, react or delete actions, including after reload. An agent's deleted answer reads "Message deleted" and adds no time. Historical progress acknowledgements stay hidden, including when their saved rows are deleted. Editing or deleting a message doesn't change a memory an agent already kept from it, but nothing more is learned from a deleted message or answer; see [Your personal memory](https://docs.connie.ai/agents/memory-and-privacy.md#your-personal-memory).
* **Add a tool from the + menu**: click **+** at the bottom left of the message box. It lists **Document**, **Mention** and **Suggestions**, with a **Search** box at the top. **Suggestions** lists tools only, one per provider, so a provider with many tools appears once; type to search every tool. Pick one and it's written into your message as **@Name**, and the agent you message works with it. You can pick only what you can open yourself, and the agent reads it with its own access, so it never sees anything through you that it couldn't open itself. If you delete the **@Name** before sending, it isn't sent. If something you picked can't be opened when the agent starts (it was deleted or its sharing changed), the agent says it's unavailable and doesn't use something else instead.
* **Attach files** and **dictate** as in chat: click **+** and choose **Document** (or paste or drop a file). Everyone in the room sees the files you send above your message; click one to open it beside the conversation. Someone who leaves the room or is removed from it can no longer open them: a link to one then says "This file is no longer available. It may have been deleted.", the same as for a file that doesn't exist, so it doesn't tell them the file is there. A file that couldn't be uploaded stays in the message box in red saying why, and the message can't be sent until it's uploaded or removed. If the connection was lost, the file says "\[file name] didn't upload because the connection was lost. Check that you're online, then click Retry.": once you're back online, click **Retry** on it to upload the same file again, without picking it again. A file still failed after you switch rooms or reload has no **Retry** and says to remove it and attach it again. See [Troubleshooting](https://docs.connie.ai/chat/overview.md#troubleshooting) in chat.

A message you send shows **Sending…** until it is saved. If it couldn't be sent it shows **Message couldn’t be sent.** with an info icon and **Retry**, which sends the same message once.

Agent messages do not show an “In reply to” preview above their content. People’s explicit replies still show the message they reply to.

Hover actions for your sent messages sit immediately before the message bubble, including when its reply preview is wider than the bubble.

The **@** mention dropdown opens directly above the **@** text with a 4px gap, even when the composer includes a connection banner, attachments or a reply preview. A selected person or agent appears in the message box as it will read once sent: the person's profile picture or initial, or the agent's avatar, then the name, in blue for a person and in the agent's own color for an agent. **@everyone** and other mentions appear in blue text with their **@**.

A message you haven't sent stays with its room or DM, with its mentions and attached files, as in chat, and with the message it replies to: "Replying to \[name]" comes back above it too, quoting that message as it reads now, unless it was deleted or is no longer in the room. If that message is deleted while you write, "Replying to" goes and your message is sent without it. Opening another room or DM shows that conversation's own unsent message (or an empty box), and yours comes back when you return, also after a page reload. Unsent messages are kept in this browser tab and are never sent on their own. They're yours only: someone else who signs in on this tab doesn't see them, signing out removes them, and one left unopened for more than a day is removed. A prompt a link puts in your DM's message box goes above an unsent message already waiting there.

A message an agent or Connie sent for someone, for example when you ask Chief in your DM to post in a room, or ask Connie in Claude or ChatGPT, is that person's message, with a gray line under their name saying who sent it: "Sent for Yann by Chief", or "Sent for Yann by Connie". On a message sent for you it says "Sent by Chief" (your own messages carry no name).

Agents can address another agent already in the room with **@Agent name** in their normal reply. A question or explicit handoff can invite that agent to continue in the same conversation. Mentioned agents show their avatar in current and saved messages when their identity is unambiguous and still visible to you, including agents outside the conversation that appear in your sidebar. An incidental, quoted or refused mention alone does not invite another reply. This grants no access to an absent agent or private conversation.

Internal agent exchanges in rooms and agent DMs show a centered summary such as **4 messages with Scout**. Click an agent's name to read the exchange in a side panel beside your conversation, replacing **Chat details** while open. The main conversation and message box stay available, and your draft stays in place. The header shows the two agents’ stacked avatars and a short AI-generated conversation title, without a name pill or exchange arrow. Names above messages keep each agent’s readable color, as in rooms. The title loads separately from the messages. A **Close agent conversation** button remains in the header; Escape while inside it also closes it. If Chat details was open, it returns when you close the exchange. Internal coordination stays in the read-only transcript. Each linked name opens a separate two-agent thread, so a broadcast recipient sees only its own exchange with the sender. Requested drafts and useful final outcomes also appear in the main conversation for your review, and the coordinating agent reports one consolidated broadcast outcome. Agents message separately by default; ask for a group when you want them in a room. Direct answers, deliberate room posts and questions or approvals that need your decision stay in the main conversation. Access to the transcript is rechecked on every read.

A confirmed send without a readable exchange summary and a deliberate room post retain the small **Messaged** line with each recipient's avatar and name outside the answer bubble. Several recipients share one line. This receipt comes from confirmed delivery, rather than the agent saying it sent something, and remains after you reopen the chat. Failed, unconfirmed and proposed sends do not show a success receipt. Click a recipient to read the authorized agent conversation in the same side panel; observing it sends no message and grants no additional access.

In a page comment, **Mention** also lets you choose agents you own and rooms you have joined and can post in. After the comment saves, it sends your comment and a page link under your name to the agent’s private conversation or the selected room. Sharing an agent with you or meeting it in a room does not let you tag it from a page. Room handoffs use the same routing described below, and agents can ask for clarification. The page link retains its access rules. See [page comments](https://docs.connie.ai/chat/sharing-and-artifacts.md#find-your-pages) for delivery and retry details.

## Who answers

If a room has no agents, including after you remove its last agent, the message box says **Message the room...**. Above it, **No agents in this room** explains: "Messages here won’t get an agent response until an agent is added." You can still message people in the room. If you can add participants, click **Add an agent** to open **Members** with the agent picker; otherwise, ask a room manager to add one. Adding an agent restores the usual message box without sending your unsent message. Paused agents still count as present because they can answer direct requests.

Only one agent normally takes a message, chosen in this order:

1. **@everyone** addresses every agent present that can answer. Otherwise, the agent already in this room that you **@mention** answers. Several mentioned agents each receive the message, unless you assign an order or dependency: then the agent responsible for the first unfinished step starts, and later participants wait for their turn. A tag naming an absent teammate or agent leaves the message with one agent here, which explains how to reach them or handles your explicit request; it does not wake an agent elsewhere.
2. The agent whose message you **reply** to.
3. The agent already working on your request.
4. The room's only agent.
5. Otherwise Connie picks the agent the message is for, or you're asked to say who you mean.

For independent assignments, put each agent's exact `@Name` and instruction in its own paragraph, for example `@Researcher: Compare the companies.` and `@Writer: Update the budget.` Clear separate assignments can reach those agents even when pasted as plain text without selecting mention suggestions. Names must uniquely identify agents already in the room; quotes, ambiguous references and ordered handoffs do not become parallel assignments. Selecting names from the mention picker remains the clearest way to identify them.

People talking to each other get no agent reply. In a room with one person and one agent, every message goes to the agent.

If you are the only person in a room with several agents, a greeting such as "hello" gets one agent's reply. With other people present, greet the agents explicitly, for example "Hi agents", or mention the agent you want to answer. Greetings between people, thanks and reactions do not start agent work.

When you ask several agents in a room to complete work together, a language model decides who should contribute next or whether the work is finished. When you describe an order of work using agents' names or individual **@mentions**, the agent assigned the first step starts; the other agents take their turns when needed. For example, “@Artifacts Manager, import these Records; once done, tell @ToolBot which tools you used” starts Artifacts Manager first. ToolBot does not repeat the import. Independent assignments can still start together. An interviewer and candidate can take turns, or a writer can hand a draft to a reviewer and revise it after feedback, without you sending a message between turns. An **@mention** helps identify someone, but does not force another reply; a tag in a quote, a thank-you or a finished result does not by itself restart the discussion. There is no fixed turn limit. Agents still wait for your answers, approvals and connections, use the workspace's normal credits and permissions, and stop when you use **Stop** in the message box.

While an agent works, its activity line shows its face and one short line at a time: the step it is running or just finished (for example **Searching the web** or **Read web pages**). That step stays visible while it runs. Before the first step, the line shows thinking words; writing words appear when no step is available. For longer work, it can also post short progress messages explaining what it completed and what comes next. A created artifact appears as soon as it is ready to open: for example, you can open a new List while the agent continues adding fields or enriching it. These progress messages and cards remain in the conversation after reload. Draft text and reasoning stay private, and tool calls are never listed in the conversation. The agent posts its final answer when it is finished. When it needs an approval, a connection or an answer from you, it posts the card with a short explanation and its activity clears while it waits.

When an enrichment or Records review finishes, its completion receipt keeps the agent that started the run as its sender, including after reopening the conversation. If its original author cannot be verified, it appears as a neutral notice.

For work that takes time, the agent can first post one short message describing what it will prepare or check, then follow up with the result or the decision it needs from you. It is one sentence of at most 25 words, and it skips generic “let me look into it” acknowledgements. It never says it will save, change, rename or delete something before the agent has tried, because you may not be allowed to make that change: it says what it will check or try, or the agent goes straight to its answer. When the reply itself is the whole job and nothing has to be looked up or saved (rewording a text you pasted, a short draft from details you gave, an explanation), there is no first message: the answer itself comes next. A question about your workspace's own data, such as how many people a List holds, does get one, because the agent has to look it up. That first message does not mean a tool has run, an account is connected or an external action has been approved. The agent knows what its first message said: if it ends up doing less or something different, its answer says so. Its activity continues while the actual work runs. Exact-output requests, control messages and scheduled/background work keep their existing reply behavior. Research and external services can still take time.

After **Stop**, stopped work stays stopped. A new message assigning work to both agents starts fresh work for both; retrying the stopped message does not restart it.

Several agents can work in the same room at once, and one agent can handle several independent tasks there or across rooms. Each running request has its own activity; one finishing or waiting does not clear the others. The activity area shows one request at a time, switching directly to the next every two seconds. Activity words also change directly, keeping the indicator visible while its avatar moves and its text shimmers. You can keep talking while they work. Reopening the room restores active work without showing unfinished replies.

When you add to your own request while an agent is working, a message it has already taken into that work stays part of the request after an approval, connection or question pause. Continuing the work keeps both the original request and those additions, along with the steps already completed. Removed messages and files the conversation can no longer read are not brought back into the continuation. The answer always appears below your latest message: if the agent had already shown you a reply for that work (for example with an approval card) before you added to it, that reply stays as you read it and the answer follows your message, saying what changed rather than repeating itself.

People typing also appear above the message box. The line names up to three people; past three it names two and counts the rest. With more than three typing it shows two faces and **+N**; hover **+N** or focus it with the keyboard to see the other names. A follow-up to your current task steers that work. Ask for a separate task when you want it to run independently. With only one task it could belong to, your message goes straight to that task. If several of your tasks could match, a **Which task is this for?** card lists them with **New task**; click one and your original message goes to it, or type its name. You're only asked once: if your answer still doesn't name a task, your original message starts as a new task. An approval, connection or question continues only its own task and can be decided while other tasks run. A task may briefly finish the step that produced its card; your submitted answer stays pending and continues automatically.

## Guidance for a room

Tell the agent in the conversation how you want it to work here, for example "Answer in French, and keep replies short." There is no **Instructions** option or editor. Room-specific requests do not change an agent's purpose in other conversations.

To keep facts or preferences for future conversations, tell one of your own agents in a private DM, where Memory is available. Reusable procedures belong in skills. Memory and skills keep their existing access rules; asking in a shared room does not make private guidance available to everyone. See [Memory and privacy](https://docs.connie.ai/agents/memory-and-privacy.md).

## Rename, duplicate, leave or delete a room

Click the centered room name pill:

* **Members**: who is in the room; see [Members](#members).
* **Add people** (non-guest participants): opens **Add people**, as the button at the top right does.
* **Add agents** (non-guest participants): opens **Members** ready to add agents.
* **Export**: downloads the room as a Markdown (`.md`) file, everything in it you can read, each message under its author's name. A deleted message reads "Message deleted" there too, with the steps of an agent's deleted answer. In an agent's DM, **Export** is in the agent's menu at the top left.
* **Rename room** (non-guest participants): type the new name and press **Enter**.
* **Duplicate** (anyone who can post): makes a new room you participate in with the same saved room guidance and access, the same people with equal room controls (guests read only) and the agents you can use, and an empty conversation (no previous messages are copied, though one of its agents may start it with a short message), then opens it ("\[room] was created"). The sidebar shows **Duplicating…** while the copy is saved; it opens only after the server confirms it. A failed copy disappears from the sidebar and leaves the original unchanged.
* **Unarchive** (non-guest participants), only on a room that was archived: rooms can no longer be archived, but an archived room stays readable and nobody can send messages in it until it is unarchived ("This room is archived. Unarchive it to send messages again."). Archived rooms are under **Archived rooms** at the end of the sidebar's list, marked **Archived**; open one there to read it.
* **Leave room**: "You will no longer see its messages. Someone in the room can add you back." The remaining participants keep the same controls. If no non-guest participant remains, the room is archived.
* **Delete room** (non-guest participants): "It is removed for everyone in it, and its scheduled work stops. Agents in it are not deleted, and anything created here keeps its own access." The routines that report into the room end with it at once: they are removed and their event triggers are disconnected. The room disappears for everyone, including from the sidebar and **Search**. Agents stop the work they were doing in it, and each person whose request they were working on gets a **\[agent]'s work stopped** notification: "The room it was asked in was deleted, so it stopped working on it." Pages made in it stay in **Artifacts** → **Pages** for the people who were in the room when it was deleted (and for everyone in the workspace if it was open to them): those who could post can still edit them, the others can read them, and a page's own sharing still applies. Files sent in the room open only where they were shared on their own, for example with a link. Apps made in it stay open in the same way as its pages, to the people who were in the room when it was deleted (and to everyone in the workspace if it was open to them); only an app's creator and the people it was shared with to edit can change it, only its creator and the people it was shared with can use its tools, and its own sharing still applies. Records, Lists and Workflows keep their own access, as always.
* **\[agent]’s details** and **Manage \[agent]’s permissions**, for each agent in the room. See [Agents and permissions](https://docs.connie.ai/agents/agents-and-permissions.md).

Someone removed from the workspace leaves all its rooms the same way: the remaining participants keep the same controls, and a room with no non-guest participant left is archived. Their messages stay.

Leaving or deleting a room never deletes an agent. Deleting an agent removes it from every room it was in; its messages stay. See [Agents and permissions](https://docs.connie.ai/agents/agents-and-permissions.md).

## Good to know

* **Click a room’s three-dot button in the sidebar** for its options: **Pin to top** and **Mute** for yourself (pinned rooms come first, a muted room's unread count turns gray); **Add people**, **Rename room** and **Delete room** for every non-guest participant; and **Duplicate**, which makes a new room with the same saved room guidance and access, the same people with equal room controls (guests read only) and the agents you can use, and an empty conversation (no messages, tasks or approvals are copied; one of its agents may start it with a short message).
* People who can only read see "You can read this room but not send messages in it." in place of the message box.
* A room's history is shared with everyone added to it, but files, Records, Lists and other artifacts linked in it keep their own access. See [Memory and privacy](https://docs.connie.ai/agents/memory-and-privacy.md).

## Related

* [Agents and rooms](https://docs.connie.ai/agents/overview.md)
* [Work in progress](https://docs.connie.ai/agents/tasks-and-work.md)
* [Members and roles](https://docs.connie.ai/workspace/members-and-roles.md)

The centered chat name opens a floating details panel beside the conversation. It shows the chat name, all room members’ names, participant avatars and member count, work in progress, and artifacts. Every artifact type in this panel uses a 16 × 16px icon with the document icon’s stroke weight (about 1.33px), including newly added artifacts. Every member’s avatar appears side by side in a single row without outlines; scroll the row horizontally if there are more avatars than fit. Close it with the close button or click the chat name again.

Question options in rooms and agent DMs show their label and any explanation together as one piece of text. Selecting an option still sends its original answer.

An open, standalone single-choice question fills the conversation column. Quote and plan questions keep their standard card width.

Multi-question cards show one dot per question at the bottom left. The darker dot marks the current question.

When a room or agent-DM draft wraps beyond one line, the text moves above the attachment, microphone and send controls. The controls stay together along the bottom; shortening the draft to one line restores the compact layout.

While an agent works, its activity appears at the next reply position in the message column. Sending a message leaves approximately 30% of the chat area for the reply; the same gap remains below the latest received message. Scrolling up keeps your reading position.

Message hover shows only the action icons. Click **React** to open the emoji picker; no quick-reaction bubble appears automatically.

In a 1:1 agent DM, your sent message bubbles use a darker shade of that agent’s color with white text, including older messages. Shared rooms keep their existing sent-message color.

The message area fades into white above the composer, mirroring the header fade without adding height to the composer area.

When messages remain below the visible chat area, a small down-arrow button appears above the composer. Click **Scroll to latest message** to return to the newest message.

The send button in a 1:1 agent DM matches the agent’s color. When inactive, it keeps the same color at reduced opacity.

Reaction badges in 1:1 agent chats match their message bubble with a slight lighter tint.

Agent mentions keep the avatar and name vertically centered with the surrounding message text.

In shared rooms, the send button uses the same black as sent-message bubbles and retains reduced opacity when inactive.

Agent chats and rooms use a loading placeholder matching the current message column and compact composer, including during navigation and initial history loading.

A blue **NEW** divider marks the first unread message received while you were away from the conversation. Replies arriving while the chat is visible do not start a divider. Opening a room or agent chat uses your own saved read position; reloading messages you have already read does not mark them new.

If new messages arrive while you are scrolled up, a **N new messages** label appears above the composer in place of the down arrow, in the sent-message color. Click it to reach the latest message, or click its X to dismiss the label without moving your reading position.

Both `/agent/` and `/room/` conversation routes use the dedicated room loading layout, without the earlier chat welcome or large composer placeholder.

Generated images appear once per reply with their download and enlargement controls. When an answer includes an artifact followed by a closing question, that question does not repeat the image above the artifact. Message drafts also keep a single generated-image preview.

Question cards following an agent’s tool-activity bubble keep the same 4px gap, including when an empty artifact wrapper sits between them.

The new-message label counts only loaded, saved messages, never Thinking, Working, pending sends or unfinished agent replies. The label has no border.

At narrow chat widths of 400px or less, messages and artifacts can fill the available message column rather than being limited to 80% of it. The outer chat padding remains.

At chat widths of 400px or less, hovering a message shows no action icons or corner dots. Right-click its bubble to open the available actions directly, including **Reply** and **React**. With the keyboard, Tab to the bubble and press **Shift+F10** or the **Menu** key; **Escape** closes the menu and returns focus to the bubble. Wider conversations keep their existing hover controls.

The round agent face uses a true circle throughout the app, including sidebars, messages, profiles and avatar previews.

Agent start-failure notices appear as muted gray status text without hover message actions.

Agent faces smile while they rest, blink now and then and, in the sidebar, lean their eyes toward your pointer. A paused agent's face sleeps. While an agent thinks or works, its face stops smiling: it opens its eyes and looks around, ponders, focuses and squeezes its eyes in turn, changing expression behind a blink. Thinking avatars look around for 1.5 seconds, play one backflip, then gently sway left and right with a longer closed-eye blink. Reduced-motion preferences disable these animations and show open eyes while the agent works.

Choose **Hexagon face** in the agent’s **Face** picker to use the rounded hexagon avatar, pointed at the sides with a flat top and bottom. It supports the same colors and working animations as the other faces.

Thinking avatars begin facing forward with centered, upright eyes before looking around.

After deciding a tool approval, the agent’s thinking indicator stays visible while the decision is processed and hands over to live activity when work continues.

While thinking, agents smoothly look in all directions. Occasional shakes, closed-eye sways, happen at random intervals rather than after a fixed sequence.

Consecutive tool-step continuations of the same answer join into one continuous rounded bubble with matching width.

Thinking avatars backflip only once, after looking around for 1.5 seconds. No later spins or side curves are shown.

In the collapsed sidebar, Search, Inbox, Artifacts, Plugins, the divider and the add button stay fixed. Only the agents and rooms list scrolls, fading at edges with more entries.

The scroll-to-latest control is 32 × 32px with an 18 × 18px down arrow.

Published tool-only continuation parts join the continued answer in one bubble. While that task continues, its activity indicator stays outside message bubbles at the end of the conversation; completed reply content joins the answer when published.

Newly published agent replies reveal text progressively inside their bubble. Historical messages stay fully visible, and reduced-motion preferences show the full reply immediately. This reveals only published replies, never unfinished drafts.

Message mentions show the person’s profile picture or initials, or the agent’s avatar, followed by a medium-weight name without **@**: blue for a person, and the agent’s own color for an agent (darkened where it would be hard to read). A 6px gap separates each complete mention from the surrounding sentence; punctuation and line breaks stay as written. In your own sent bubbles, agent names use a nearly white tint of the agent’s color and people’s names keep the bubble’s white text. In both your messages and agent replies, hover an agent mention to highlight its chip and see its name, or focus it with the keyboard. Click it, or press Enter or Space, to open your private conversation with that agent. Opening a mention checks that you can still use the agent; it does not share the original conversation or send a message.

When an agent links to a room you can see in your sidebar, the reference shows a room icon and its current name with the same hover and keyboard treatment. Click it, or press Enter or Space, to open that room in the current tab. Saved replies and older room addresses use this treatment too. Opening a reference still checks your room access and sends no message. A room that cannot be identified from your current sidebar keeps its original link; a room reference does not invite anyone or grant access.

When an agent DM or room has a task waiting for a human answer, approval or account connection, its sidebar preview starts with **Waiting for response**, followed by the required answer, approval or account connection, with amber (#BD740B) text at weight 450 and a matching dot. The normal message preview returns once nothing is waiting.

The blue unread dot marks unread conversations; sidebar previews keep their existing text appearance. When work stops or its activity label clears, the saved last-message preview returns.

The **New agent / New room** menu keeps its neutral border when hovered or focused; keyboard focus remains visible on its menu items.

Starter suggestions use the same Clarification card as agent questions: lettered choices, the shared single-selection style, and an inline **Something else** answer. Clicking a choice sends it immediately as a new chat message. For your own answer, type in **Something else**, then tap **Submit** or press Enter. **Submit** appears when you type an answer and disappears if you clear it.

Sidebar three-dot menus open above their trigger when there is insufficient room below, keeping the menu within view near the bottom of the sidebar.

Waiting previews truncate to one line. Hover to read the full required action, including the account provider when the saved connection card identifies it.

Sent and received reaction badges share 28px height, 4px vertical and 6px horizontal padding, and 16px emoji, retaining their respective bubble colors. Multiple-reaction counts can widen the badge.

To rename an agent you own, open its sidebar options and choose **Rename agent**. The name becomes editable directly in the sidebar, without a dialog. Type the new **Agent name**, then press **Enter** or click away to save. Press **Escape** to cancel. Blank or unchanged names keep the original name. If saving fails, the draft stays in place so you can retry. Room participants who can post can choose **Rename room** in the room’s sidebar options.

### Collaborating in a call

Ask agents to work together toward one result, or give them explicitly separate
assignments. For shared work, one agent gathers the facts, the others review
what matters, and the lead posts the final result. You can clarify the scope
aloud while work continues; a status check does not restart it.

You can ask an agent to invite another eligible agent into the call. Connie
checks normal sharing and room permissions before adding them. Agents can leave
when their part is finished or when you ask them to leave. Ending audio keeps
already assigned work in the room so you can read the result later. Tool
approvals and account requirements still apply.

Each ended call is saved as a Page in the room's details, with an AI summary
above the transcript, so you can revisit the conversation and decisions. A
transcript preserves recognized words; it is not a recording of all audio.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.