Chat

Chat is where you work directly with XTM One in natural language. It supports both the platform-wide General Assistant and individual agents.

What Chat is for

Use Chat when you want to:

ask a question in plain language
work with a specific assistant
upload supporting files
reuse a saved prompt
continue an earlier conversation
speak instead of typing, if your browser supports it

The chat layout

Chat has two main areas:

the left sidebar for assistant selection and conversation history
the main panel for the active conversation

If you are on a smaller screen, the sidebar can collapse so the conversation has more space.

Choose who you want to talk to

At the top of the sidebar, you choose the assistant for the conversation.

You may see:

General Assistant
chat-enabled agents
assignment-oriented agents shown in a Not recommended for chat section

Choose the General Assistant when you want a broad helper. Choose a specific agent when you want an assistant with a focused role or built-in context.

Start a new conversation

To start a new conversation:

Open Chat from the sidebar.
Choose the assistant you want.
Click the New conversation button.
Type your message and send it.

If you open Chat from an agent card or an agent detail page, XTM One can preselect that agent for you.

The sidebar is not only a list of past chats. It is also your main control area.

From here you can:

search conversations by title, preview text, or agent name
open an earlier conversation
organize conversations in workspaces (see below)
archive a single conversation
archive all conversations from the Clear all conversations action
switch assistants without leaving the page

Archiving removes a conversation from the active list. It is not the same as simply closing the page.

Organize conversations in workspaces

To keep the conversations of one topic together, and let their agents share work items and shared state, file them in a workspace. The sidebar lists your conversations by workspace, a little like folders:

Workspaces lists each workspace you can file into, your default first, with the conversations filed in it. Click a workspace's name to fold or unfold it; the sidebar remembers what you folded.
Not in a workspace lists the conversations filed nowhere. That is a normal place for a conversation to be: a new conversation starts there, and it stays there until you file it.

From the sidebar you can:

create a workspace with the New workspace button next to Workspaces: type its name and press Enter
file a conversation by dragging it onto a workspace, or with the Move to workspace menu of its row, which also offers Remove from workspace; dropping a conversation on Not in a workspace takes it out of its workspace too
start a conversation directly in a workspace with the + button of that workspace (New conversation in this workspace)
manage a workspace from its Workspace actions menu: Rename, and the same actions as on the Workspaces page — Make default, Knowledge bases, Share, Archive and Delete — plus Open in Workspaces. Archiving asks for a confirmation, then takes the workspace out of the sidebar; its conversations are then listed under Other workspaces. To restore it, open Work > Operations > Workspaces, click Show archived, then choose Restore in its actions menu. Deleting asks for a confirmation and keeps its conversations, which then appear under Not in a workspace; a workspace that still holds work items cannot be deleted

Only your own conversations can be filed, into workspaces you can change. A workspace shared with you in read-only shows only while it holds one of your conversations, and a conversation filed in a workspace you can no longer see is listed under Other workspaces. While you search, only the workspaces holding a matching conversation are shown.

You can also file the conversation you are in from the Conversation options menu at the top of the conversation (File in a workspace). An assistant with the XTM One built-in tool can do it for you — switch the tool on in the session tools panel, then ask for example "create a workspace called Acme audit and move this conversation into it".

Write your message

The composer at the bottom of the screen supports several input methods.

You can:

type a normal message
attach files with the paperclip button
paste files into the composer
drag and drop files onto the page
insert a saved prompt template
use voice dictation
start a full voice conversation

If a response is still being generated, the send button becomes a stop button so you can interrupt the response.

If a response ends before its answer because XTM One restarted or the reply failed, the work the agent already did is kept: the next message you send in the conversation (Go on, for instance) gets the tools it called and what they returned, so the agent continues from there instead of starting over. A response you stop yourself, and one that ended more than 24 hours ago, start from scratch. A long piece of reasoning with no tool call cannot be kept.

A line under the composer reminds you that AI can make mistakes and that important information should be checked. The same notice appears at the bottom of the embeddable chat widget, between the input box and the "Powered by" line. It is shown in your interface language in the app and in English in the widget.

The lightning icon among the composer buttons counts the agentic actions consumed in the current quota period (the day, month or year your administrator configured), out of your limit when there is one. Hover it for the details. While an assistant works on your message, the action this reply consumes is shown in brackets: 4(+1)/10 means 4 actions before this reply, 1 for it, out of 10. It joins the total when the reply ends, along with any sandbox code runs counted toward the same limit; a reply that fails may have its action credited back instead (see Configuration). The counter is refreshed as well when a reply is stopped or fails, and when a reply you left running by opening another conversation ends. A message you send while a reply is running can count as one more action.

Ask several agents at once with @mentions

Type @ in the composer and pick an agent from the list to mention it. You can mention several agents in the same message.

In a new conversation, the first agent you mention becomes the conversation's agent.
The conversation's agent answers first. Then every other agent you mentioned answers in turn, in the order you mentioned them. Each reply is its own message, labelled with the agent's name.
Every mentioned agent receives your full message, including attachments, and can read the replies given before its own in the same turn.
Each agent is told which agents answer after it in the turn, so it answers its own part and does not ask you to pass your message on to them.
An agent you mention twice answers once.
Only agents you can chat with answer: a mention of an agent that is not available to you (not shared with you, disabled, or not open to chat) is ignored, and the others still answer. If an agent fails while answering, a short note replaces its reply and the agents after it still answer.
Clicking stop ends the turn: the agents that have not answered yet are not asked.
The mentioned agents of one message share a time limit. An agent still answering when it runs out is stopped, and a note names the agents that were not asked.
Each agent that answers counts as one agentic action, charged before it starts, exactly as if you had asked it alone; an agent it consults while answering adds nothing. If your agentic quota runs out, the agent that could not be charged says so and the agents after it are not asked. An agent that fails is credited back like any failed answer.
Each reply keeps what its own agent produced: its trace (for feedback) and the files it generated.
A message you send while an answer is still being written is held and sent as its own message when the answer ends if it mentions another agent, or while several mentioned agents are answering. A message that reaches a running answer anyway (from another device, or through the API) goes to the agent already answering, and the other agents it mentions then answer it, after the agents of the running turn and under the same rules. One that arrives as the answer is ending is answered by them after the next answer in the conversation.

Once an administrator turns them on (CHANNEL_AGENT_MENTIONS_ENABLED=true, see Configuration), mentions also work in messaging channels (Slack, Teams, WhatsApp, Telegram, Discord and Google Chat), for people with an XTM One account. Until then, a channel message is answered by its routed agent alone, whatever it mentions. There is no picker there: type the agent's key after @, the same key the web composer inserts, such as @mail-desk. A mentioned agent answers only when its own channel settings let it answer that message: it is enabled, open to chat, set to Receive on that channel (and on that channel conversation, if it is restricted to some), past the bot's receive whitelist, the private-conversation and only-mention rules, and set to Publish there. With mentions turned on, an administrator decides which agents can answer on a channel in each agent's Channels settings. Any other @ word - a person, a handle, an agent not set up on that channel - is plain text: the routed agent answers alone, as it would without the mention, and no note is added. Anonymous senders cannot mention agents. Each agent that answers counts as one agentic action, as in the web chat, and is told which agents answer after it. When the routed agent has nothing to add, the agents you named still answer. The replies arrive together in one bot message, one section per agent, with a link to every file an agent produced, and are stored as separate messages in the conversation, each with its own agent's files.

Point the agent at another conversation with @

To have the agent read another conversation - "@q3-threat-report, what did we decide there?" - type @ in the composer: below the agents, Reference a conversation lists the conversations you can open, the most recent first. Type words of a title to narrow the list, joined with hyphens (@q3-threat finds "Q3 threat report"), then pick one: the composer inserts its key, the title in lowercase with hyphens, such as @q3-threat-report.

The list holds your own conversations, from the web chat and from the assistant embedded in OpenCTI or OpenAEV, and those shared with you by name. It leaves out archived and empty conversations, the conversation you are writing in, and conversations only reachable through an organisation link.
Only a conversation you picked from the list counts: an @ word you type by hand stays plain text. Deleting the key from your message removes the reference.
With your message the agent receives a short brief of each conversation you referenced: its title, the summary and the instructions kept from its earlier messages, and its last few messages. Every agent that answers the message reads it, the mentioned ones included.
For the rest of the conversation, the agent can read the referenced conversations in full, page by page. It can read only the conversations referenced in the conversation it answers, and only while you can still open them: one deleted or no longer shared with you is no longer readable.
Your message shows a chip for each conversation it referenced. Clicking it opens that conversation. Someone you share your conversation with sees the chips, which name the referenced conversations, but can open only the ones they have access to.
A message references at most 5 conversations.
A message that references a conversation while an answer is still being written is held and sent when the answer ends: a correction sent to a running answer cannot carry a reference.

The Presentation Studio's composer offers the same list, with conversations only. In the assistant embedded in OpenCTI or OpenAEV, the list appears once the platform enables it.

Ask for changes to a file the agent produced

When an agent hands you a file — an HTML briefing, a CSV export, a Markdown report, a JSON bundle — you can ask for a change in plain words: "remove the paragraph about X", "rename the column to Y", "fix the date in the footer". The agent reads the file it produced, applies exactly the change you asked for and returns a new revision of the file. The version you already downloaded is never altered under you.

Text-based files (HTML, Markdown, CSV, JSON, YAML, source code, ...) are edited in place, so a one-line change stays a one-line change — the rest of the document is untouched.
Rendered documents (PDF, Word) have no editable source: the agent regenerates them from the corrected content.
Files you uploaded yourself can be corrected the same way; the result is a new file next to your original.

Tip

Name what you want changed as it appears in the document. The agent matches your words against the file's exact text, so "remove the sentence about JP" works when the document says "JP"; a paraphrase may need one round of clarification.

How a reply shows its files

The files an answer names are shown as cards in the answer. The other files the agent wrote while working on it (a first draft, a scratch file) are listed under a collapsed Working files row below it, still downloadable.
An answer that names none of its files shows them all as cards, as before.
When the agent wrote a file twice under the same name in one answer, the older one is marked Superseded and Go to the newer version takes you to the newer one.
Files that share a name and type in a conversation are numbered v1, v2, ... in the order they were written, and their cards show the exact size in bytes. Every card shows the time its file was written; hover a card to see its file id and the start of its SHA-256 digest.

Use prompt templates

If prompts are available in your environment, the prompt button opens a list of reusable prompt templates.

Use this when you want to:

start from a standard request
keep your wording consistent
save time on recurring questions

Selecting a prompt inserts its content into your message box.

Use session tools

The wrench button opens the session tools panel.

Depending on the conversation, you may be able to enable or disable:

built-in tools
custom tools
integrations
MCP servers
knowledge bases

When you use the General Assistant, this panel lets you build a temporary toolset for the conversation. When you use a specific agent, the panel starts from that agent's own toolset and lets you switch built-in tools on or off for the conversation. Work items, shared state and trace inspection act as the agent itself, so they are only offered when the agent already has them or when you can configure the agent. Trace inspection reads the agent's traces, which hold every user's conversations with it, so the one you add, and the one the platform builder (the suite named after the platform) brings along, only work for someone who can read the agent's chats: an admin, its owner, or someone it is shared with as Full access (incl. chats).

A message can only add to the conversation what you could connect to an agent of your own: one that adds a custom tool, MCP server, knowledge base or integration you cannot see is refused with an error that says which kind it is, and nothing is sent. Saving your profile's default tools follows the same rule. What the conversation already has - the tools you selected earlier, your profile defaults, the agent's own toolset and the General Assistant's configuration - is used as it was set up, including what the owner of an agent shared with you connected to it.

Integrations are the one exception, as they always were: an integration you cannot see, or whose owner's connected account is not shared with you, is never used for you in a chat (see Sharing an OAuth integration). Nothing is hidden: when an answer could not use one, a note under it names the integration - by its service only when it is not shared with you - and says why, and the agent is told as well, so it can say what it cannot do. Connect your own integration of that service, or ask the owner to share theirs.

The panel also lets you:

reset back to the default configuration
enable all visible tools (the ones already on stay on)
disable all visible tools
temporarily hide integrations from the picker and restore them later

Voice options

If your browser supports speech features, Chat can offer two voice modes.

Voice input converts your speech into text in the message box.
Voice conversation opens a more continuous spoken interaction.

If these controls do not appear, your browser or device does not currently support them.

What you may see above the messages

The conversation header can show extra context.

You may see:

the active agent name
the current channel, such as web chat
a transfer banner if another agent handed the conversation over
a Return action if you want to go back to the source agent

This is especially useful when the conversation moved from one agent to another.

A conversation open in several places

When the same conversation is open in several tabs, on another device, or in voice mode, each one shows the answer as it is written, wherever the message was sent from. Voice mode reads out only the answer to what you said.

When XTM One is updated while a tab stays open, that tab shows A new version is available: select Reload to use the new version. Until you do, the tab keeps working, but it may show an answer started elsewhere as if you had asked it.

To connect your own scripts to a conversation, see the Chat API.

Continue a conversation in Slack or Teams, or back on the web

A conversation is not tied to the place it started. The same conversation - its history, files, session tools and agent - can go on in the web app, in a direct message with the Slack bot, or in your personal chat with the Microsoft Teams bot.

From Slack or Teams to the web. A direct-message conversation you had with the bot appears in your conversation list, marked with the Slack or Teams icon. Open it and write: the agent answers here, with everything said in Slack or Teams. Each message in the transcript shows where it was written.
From the web to Slack or Teams. Open the conversation, then Conversation options > Continue in Slack or Teams, and pick the app. The bot sends you a direct message: in Slack, reply in its thread; in Teams, reply in your chat with the bot. The conversation goes on there, and stays in your list here.
Each reply is written where you send your message, with that app's formatting. A message you send in Slack is answered in Slack; one you send here is answered here.
The same agent answers everywhere. When your administrator linked the General Assistant to an agent (Settings > General Assistant > Link to Agent), a Slack or Teams conversation with the General Assistant is answered by that agent there and here alike: the chat shows its name, and its tools in the composer. A conversation you started here with the General Assistant before the link keeps the General Assistant in Slack or Teams too.
One answer at a time. A message you send in Slack or Teams while another reply in the conversation is being written (here or there) is queued: the bot tells you so right away, and answers it as soon as that reply is done. Several messages you send meanwhile in the same thread or chat are answered together, in one reply, as the web does with a message sent while the agent is still working. Up to five messages can wait; past that the bot asks you to send yours again later. A queued message the bot could not answer - within an hour, or because something failed twice - is dropped, and the bot tells you so in the thread so you can send it again. A queued message keeps the files you sent with it. If the thread stops continuing the conversation before a queued message runs (you send /new or stop the continuation), that message is not answered, and the bot says so. While a reply to a message sent from Slack or Teams is being written, the web refuses a new message and asks you to send it again once the reply arrives.
Tool calls waiting for your approval are answered in the web app. Slack and Teams cannot show the proposal. While the turn that proposed the call still waits for your decision, a message you send in Slack or Teams is queued behind it. If that turn ended with the call still waiting, a new message moves the conversation on, as it does on the web: the call is dropped, and the reply starts by saying so.

A few limits:

Only the person who owns the conversation can continue it. A conversation shared with you stays read-only, and a conversation from someone who wrote to the bot without an XTM One account is never listed.
Group threads stay where they are. A Slack channel thread or a Teams group chat or channel holds other people's messages, and everyone in it reads the agent's answers, so it is not listed here and is continued in Slack or Teams only: you cannot send a message or upload a file into one from here.
In Teams, your personal chat with the bot has no threads: once you continue a conversation there, every message you send in that chat continues it, until you send /new or continue another conversation there. Teams also needs you to have written to the bot once, so the bot can write back to you.
The option is offered only where the bot would answer you: the conversation's agent must answer on Slack or Teams (a conversation with the General Assistant always can), and the bot must take direct messages. A Slack bot an administrator restricted to some channels with direct messages turned off does not, and the dialog says so.
Stop in the same dialog ends the continuation: the next message you send in that Slack thread or Teams chat starts a new conversation.

Export a conversation

Conversation options > Export as Markdown or Export as PDF downloads the conversation as a document: every message in order, under who wrote it and when. A chart an answer shows (see Platform insights) is written as a table of its values. Anyone who can read the conversation can export it, a shared one included; the 2,000 most recent messages are exported.

Background task banners

Some requests continue in the background.

When that happens, Chat can show:

a banner while background tasks are still running
a completion banner when those background tasks finish
a failed count if some delegated tasks did not complete successfully

This means the request involved delegation or longer-running work. It is normal behavior, not an error by itself.

Files a delegated task produces (a CSV export, a spreadsheet, a report) belong to the conversation: they appear as download chips on the message the agent posts once the task finishes, and the agent itself can open, check and merge them with its own file tools when it summarizes the result.

Work the agent hands to the background in the Video Studio (a render, an AI shot, a recording, AI music) is shown the same way once the reply ends: a banner says The assistant resumes this conversation when this work finishes and lists each job with how far along it is. When the work ends, the agent comes back on its own and tells you how it went.

When a reply stops before it finishes

A reply can stop before its end: XTM One was restarted or updated while the agent was working, or the process running it stopped. The conversation never just goes quiet when that happens.

At the end of the conversation, a notice says This response stopped before it finished, and when. It stays there after a reload and on your other devices, until a new reply starts.
Resume sends Continue where you left off. for you. The tool calls the agent had already completed are kept (the notice says how many): the agent continues from their results instead of running them again. Sending any other message continues the same way.
Dismiss hides the notice.

When XTM One restarts or is updated normally, a reply still running when the server stops waiting for it also leaves a message of its own. It says what stopped the server process running it: a deployment, a scale-in or a restart of the platform, or the recycling of a web worker that served its quota of requests (which waits for the replies running in it, for hours if need be, so it rarely cuts one). The message then lists the tool calls that had completed (a failed call is marked), the ones still running when it stopped, and the last completed action. Its reasoning details (the warning icon next to the message) show each completed call with what it was given and what it returned. The agent's memory records the interrupted request and its last completed action, in the memory the agent has switched on. A reply the agent was writing after background work finished is given the same time to end, and leaves the notice when it does not. A process that crashes or is killed leaves only the notice.

If XTM One cannot be reached while you are on the page, the chat keeps trying to reconnect for as long as the page is open, and follows the agent again as soon as it is back. A message XTM One refuses before any reply starts (you sent too many messages in a minute, for example) is never dropped silently: a Your message was not sent notice says why.

The waiting arcade

While an agent works through a longer request, a small Space Invader appears under the conversation and shoots the waiting message away, letter by letter. It is more than an animation: click it, or its Take control button, and a full Space Invaders game opens over the whole interface. When the agent is showing its reasoning instead, the gamepad button in the corner of the reasoning window opens the same game. You can also start a game at any time, no wait needed: Play Space Invaders in your user menu (top right), or the command of the same name in the command palette.

The chat embedded in OpenCTI, OpenAEV and OpenCRQ invites you here too: while its agent works, its Play Space Invaders in XTM One link opens <your XTM One address>/chat?arcade=play in a new tab, and the game starts as soon as you are signed in - even if you have to sign in first, or wait for an administrator to accept the platform's Terms of Use, as long as that happens within ten minutes of the click.

Controls: arrow keys or A/D (Q/D on an AZERTY keyboard) to move, Space to fire (hold it for autofire), right-click or X to send a missile where you aim, P to pause, M to mute, Esc to leave (the first press pauses, the second ends the game and saves it). On a touch screen, drag to move and fire, and tap the rocket for a missile.
Waves and bosses: every wave marches faster than the last, and every fifth wave brings a boss. Falling capsules hold power-ups: rapid fire, spread shot, laser, shield, extra missiles, and a few rarer ones.
The interface is fair game: your shots and the invaders' leave holes in whatever they hit - buttons, messages, the sidebar. Nothing is really broken: the damage is cosmetic, and everything is rebuilt the moment you leave. Your conversation keeps running underneath.
Your answer comes first: when the agent has finished answering, a banner says so - a message it posts along the way before carrying on with its work does not count. Esc (or Read it) takes you straight back to it, and if you had stopped playing, the arcade takes you back by itself after a few seconds. Keep playing lets you finish your game. The game waits for the conversation you started it from: if you switch to another conversation or leave the chat while you play, the game goes on but no longer announces that answer.

Leaderboard

Every game you play is scored and saved. The leaderboard - Arcade leaderboard in your user menu, the trophy button next to the waiting game, the Open the leaderboard command, or the pause menu - ranks the best game of each player over the last 24 hours, the last 7 days and all time, shows who is playing right now, and names the holders of three titles: top vandal, sharpshooter and marathon. During a game, the score bar shows whose score you are about to beat.

Your callsign is three letters or digits, like on a real arcade cabinet. It defaults to your initials; change it in the leaderboard's My arcade tab. Other players only ever see your callsign - never your name or your email.
Turn off Show me on the public leaderboard to keep your scores, titles and live games to yourself, from that moment on - even in a game already running. You still see your own rank.
Scores are checked by the server: a score, a number of shots or a count of destroyed interface elements that is not plausible for the time played is refused, and so is a score higher than the waves you reached can yield. The arcade tells you why a score was not recorded. A game in which a cheat code was used is saved but not ranked.

Achievements and secrets

The leaderboard's Achievements tab lists what you have unlocked. Many achievements are secret until you find them: old cheat codes, a very rare saucer, certain dates on the calendar, and a few things you probably should not shoot. Explore.

When the game is not offered

The gamepad always means play: on the waiting panel and in the reasoning window it opens the game, and nothing in the chat hides it. When your system asks for reduced motion, the waiting panel shows a plain rotating message instead of the invader and the chat offers no game; Play Space Invaders (in your user menu or the command palette) and a ?arcade=play link still start one, since you asked for it. The arcade is not offered while an administrator impersonates a user: nothing is ever recorded under someone else's name. Nor is it offered while the platform's Terms of Use are waiting to be accepted. Administrators can turn the arcade off for the whole platform with ARCADE_ENABLED (see Configuration).

The game and the leaderboard load in the background shortly after you sign in, so a page left open while XTM One is updated can still play. If they could not load before the update (on a Data Saver or 2G connection, for instance), opening them shows Reload to play or Reload to open the leaderboard: Reload reloads the page and opens the game or the leaderboard right away.

When an agent is shown as not recommended for chat

Some agents are built mainly for background execution rather than direct conversation.

These agents still appear in the picker, but XTM One marks them so you know they are not optimized for normal chat use.

You can still try them, but the best experience usually comes from:

the General Assistant
a chat-enabled agent

Practical tips

Start with the General Assistant if you are unsure which agent to use.
Use a dedicated agent when the topic belongs to a specific workflow.
Attach documents when the answer depends on exact wording or evidence.
Open the same conversation again if you want XTM One to keep the discussion context.

Next step

The next chapter explains Agents, including what the list page shows and how to read an agent detail page.