Live Conversations

The Supervisor section shows every agent conversation your organization has right now, across both channels: phone calls and website chats. For each one you can see which agent is handling it, who is on the other end, and how long it has been going — and you can open any of them and read the conversation as it happens.

It updates as conversations start, move, and end. You don't refresh it, and you don't pick a date range. The board is meant to be left open on a second screen.

If your organization has no agents yet, the empty board shows a Create Agent button that takes you to the Agents page to set one up.

What a row tells you

The live conversations are listed in the left-hand navigation sidebar — they are the Supervisor sidebar, so they are there whenever the section is, and the page itself shows the transcripts you have opened. Each row shows:

Above the list, a search box narrows it by agent name, phone number, or widget name, and the button beside it holds three more controls:

• Sort by — Longest (the default), Newest, or Agent A-Z. Longest leads because a conversation that has gone unusually long without moving is normally the reason to be here.
• Agent — show only one agent's conversations. Only agents that currently have one are offered, so the list empties out as they finish.
• Supervision — show conversations by who has them. See Monitoring and viewing a conversation.

The button is highlighted whenever any of them is off its default, and hovering it says how the list is currently sorted and filtered. Under the search box, a count reads 2 active · 14 idle — or 3 of 12 while search or a filter is hiding rows, so a narrowed list is never mistaken for a quiet queue.

Why chats behave differently from calls

A phone call is a connection that is either up or not. Its row exists for exactly as long as the call, and vanishes the moment the caller hangs up.

A website chat has no equivalent. A visitor sends a message, gets an answer, then reads it — and may reply in ten seconds, in ten minutes, or never. There is no moment at which they "hang up". So a chat conversation stays on this board while it is still going. It leaves in two cases:

• It ended — your agent finished it, or it reached its message limit. It leaves the board straight away.
• Nothing has been said for 30 minutes. It leaves the board, because at that point the visitor has closed the tab.

Leaving the board is not the same as being over. A visitor can still pick a conversation up for 24 hours from their last message, and if they reply it comes straight back to the board as Active — it is live again, and you should see it. After 24 hours, replying starts a new conversation instead.

Three different windows, answering three different questions:

The first is deliberately much shorter than the others: a board showing every conversation your agent could still answer would be mostly conversations nobody is having.

That's also why a chat's Duration counts from its last message rather than from when it opened. A visitor who chatted this morning and left the tab open would otherwise show up as a nineteen-hour conversation, which is true and tells you nothing. It is also what the list sorts on, so a chat nobody has touched since breakfast never sits above a call that is ringing now.

Reading a conversation as it happens

Click a conversation in the sidebar to open its transcript on the page. It fills in with everything said so far and then adds each message as it arrives — the caller's or visitor's words on one side, the agent's on the other. Picking one from another part of the app takes you to the Supervisor page on the way.

Both channels open the same way. A call streams as the words are spoken; a chat updates as messages are sent, which for an idle conversation means it sits still until the visitor comes back — the transcript you are looking at is the whole conversation, not a snapshot of it.

You can open several at once. Click more rows and they sit side by side, which is the point: this page is for watching, and comparing two conversations is usually more useful than staring at one. Up to six can be open together. Once six are open the remaining rows are unavailable until you close one — click an open row again, or the ✕ on its column.

Scrolling up in a transcript stops it from jumping to each new turn, so you can read back without the view jumping under you. Jump to Latest returns you to the bottom and resumes following.

Collapsing the navigation sidebar hides the conversation list, so anything you have open stays open and the transcripts get the full width. A single Live Conversations button stays in the collapsed rail, so the page is still one click away.

Every message carries the time it was recorded, under the message itself, with seconds, so you can see how long the agent took to answer, not just what it said. Hover a message for its full date and time, which is what you want on a chat that has been open since yesterday.

On a call the time is when the platform received the turn from the agent, so it trails the moment the caller actually spoke by however long the agent took to think. On a chat it is when each message was recorded, so a visitor's message and the reply to it carry the times they landed.

A few things to know about a transcript before you rely on it:

• A tool call appears in the transcript. Looking something up, calling out to another system, capturing keypad digits — each one gets a row of its own, where it happened, carrying how it turned out and the arguments the agent sent. Admins can expand a row to see the request and response behind it. See What the agent did.
• A call's transcript is not the recording, and it is not a legal record. Speech recognition makes mistakes, particularly on names, numbers, and accents. Read it as a good indication of what is being said, not as a verbatim quote. A chat is what was typed.
• A call's transcript is not saved here. Closing the column, leaving the page, or reloading discards it; for one you can come back to, see the agent's Recent Calls after the call finishes. A chat's is a different matter. It stays with the conversation, so re-opening it shows the same messages again for as long as the conversation is on the board.
• When a call ends its row leaves the list, but its transcript stays on screen and is marked Ended so you can finish reading. Close it when you're done. A chat your agent finished is marked Ended the same way. A chat that has only gone quiet is marked Idle instead: it hasn't ended, and the visitor may still reply.
• A call may show "Wrapping Up" for a few seconds after the caller hangs up. If the agent has caller memory switched on, it writes down what it learned about the caller once the call is over. The column reports that as it happens and then says how it went — see After the call: caller memory.
• On a very long conversation the transcript stops growing. There is a size limit; once it is reached the column says so, and the conversation carries on unaffected.

Monitoring and viewing a conversation

One supervisor at a time monitors a conversation. Opening a transcript that nobody is monitoring makes you its monitor, and its header shows a purple Monitoring badge. Anyone who opens it while you are monitoring it is viewing it: their header shows a gray View Only badge, with a line under it naming who is monitoring.

Only the supervisor monitoring a conversation can act on it: coach the agent, join, listen to the call, take it over, act on a held chat reply, or dismiss an alert. Someone viewing it can read the transcript, the coaching notes, and the alerts, and nothing else.

You stop monitoring a conversation when you close its transcript or leave the Supervisor page, and the same goes for viewing one. Monitoring does not pass to the next person by itself. A supervisor still viewing the conversation sees Nobody is monitoring this conversation and a Monitor button, so taking it on is their choice, and a dropped connection reconnects them as a viewer. To step back while keeping a transcript open, use the button with a crossed-out eye in its header (Stop monitoring — keep viewing, read-only on hover).

Joining holds a conversation for longer. One you have joined stays yours until you close it out, for up to 12 hours, whether or not its transcript is open.

In the list, conversations someone else has are grouped under Monitored by others, below the rest under Available. Their row names that supervisor and what they are doing: Monitoring in plain text, with a headphones icon while they are listening, or a Joined or On call badge once they have joined. An eye with a number counts who else is viewing, and hovering it lists their names. The sort and filter button's Supervision filter has four options: All, Unattended (nobody is monitoring it or has joined it), Mine, and Other supervisors.

Coaching the agent

Coach at the foot of a transcript opens a box for a note to the agent. The agent reads it before its next reply and never mentions it. The customer never sees it. The note appears in the transcript where it was written, marked will shape the next reply until the agent has replied with it, then applied to the reply below.

To coach on something said earlier, hover a message from the agent or the customer and click the coach icon beside it. The box opens with that message quoted above it, and the message is outlined in the transcript. The agent gets the quote with your note, so a note such as "too formal" reaches it with the reply you meant. Click another message to quote that one instead, or the × beside the quote to send an ordinary note. A quote longer than 300 characters is shortened.

What the agent did

When the agent uses a tool — looking a caller up in your system, capturing keypad digits, transferring — that call appears in the transcript as a row of its own, in the place it happened. Collapsed, the row says what the tool was, how it turned out, and how long it took when the platform timed it. Expand it and you get the arguments the agent sent, the tool's own name as your configuration spells it, and, when the tool failed, the reason it gave.

A built-in tool is labelled in plain English — Capture keypad digits, Transfer to an outside number. A tool you configured carries its own name, tidied up for reading. Either way the raw name is in the expanded row, so a row can always be matched back to the agent's configuration.

No result is not a failure, and the two are kept apart deliberately. A tool that never came back may well have succeeded a moment after the caller hung up; nobody was left to record the answer. Reading it as a failure would assert something no one knows. It is worth noticing on a transfer or a payment, and unremarkable on a call the caller ended mid-sentence.

A missing duration is normal and says nothing about the call. Tools the agent runs inside its own reasoning — ending the call, reading your knowledge base, handing back to a workflow — never pass through the part of the platform that times a tool, so they carry no duration at all. Neither does any tool on a website chat: a chat keeps nothing running between messages, so there is no clock to read. The row still reports its outcome.

Chats differ in one more way. Tool rows appear on a chat while it is live on this board, but a finished chat's transcript under Recent Chats stays text-only. Live calls, live chats, a finished call's transcript in Recent Calls, and a preview call all show them.

Request and response details

When an integration tool gets an answer from your endpoint, the row also shows the HTTP status as a small chip beside the outcome. A tool whose endpoint answered 404 reads Succeeded with a 404 chip, because the agent got an answer and carried on with it. A DNS failure or a refused connection reads 502, and an integration whose own request timeout fires first reads 408. A tool that got no answer before the tool's time limit (15 seconds by default) reads Failed and has no chip.

Everyone who can open the transcript sees the chip. Admins also get tabs in the expanded row:

A few details about what the tabs hold:

• Credentials are not in the Request tab. The integration's stored credentials are added after the request is built, so the tab shows the request without them. A {{secrets.*}} value in a header or body shows as the token, not its value.
• The Response tab is a scrubbed copy. Secret values stored on the integration are removed from the headers and body before they are shown, so it is not a byte-for-byte record of what came over the wire. On a timeout it says No response and gives the tool's error text.
• Bodies are cut at 64 KB. A longer request or response body shows Truncated at 64 KB.
• Tools that are not integrations have no request or response. Their expanded row shows Arguments and, for admins, Agent saw.
• Calls placed before this existed have no Request or Response tab. Admins still see Arguments and Agent saw, because both come from the stored transcript.

Details are for admins because a response body is your endpoint's customer data. The check is the administrator role, and the platform applies it before sending anything, so a non-admin's browser never receives the request or response. See Admin for what that gate does and does not keep out.

Details on a call or chat are kept with the transcript for 90 days and deleted with it, including on an erasure. Details on a test result are kept for 90 days. They appear on the live board for calls and chats, on a finished call under Recent Calls, in a preview call, and in the transcript of a test result.

Seeing that a visitor is typing

On a website chat, a bubble of moving dots appears at the bottom of the transcript while the visitor is composing a message — the same signal they see from your agent, pointed the other way. It disappears when their message lands.

Three things to know before you read anything into it:

• It is chat only. A phone call has no equivalent, and one is not missing: a caller is either speaking or not, and what they say arrives as a line of transcript.
• It never appears before a visitor's first message. A conversation reaches this page when its first message is answered, so there is nothing on the board to show dots against until then.
• It goes away on its own. A visitor who starts a sentence and then closes the tab, loses their connection, or walks away stops being shown as typing within a few seconds. So dots mean someone was typing very recently — not that a message is guaranteed to arrive.

It is also not a keystroke feed: the widget reports "still typing" every few seconds rather than per character, and nothing about what they are writing reaches this page until they send it.

After the call: caller memory

If the agent has caller memory switched on, the last thing it does on a call happens after the caller has already hung up: it writes down what it learned, so the caller's next call can pick up where this one left off. That takes a few seconds, and the column shows it rather than jumping straight to Ended — a call badged Wrapping Up has no one on the line, but the agent has not finished.

You will see one of these:

The first three are routine and need nothing from you. Memory not saved and Memory not confirmed are worth noting: a single one is usually a transient fault, but several across different calls point at a problem worth raising.

A call on an agent with caller memory switched off shows none of this and goes straight to Ended.

Taking a chat over

A chat tile answered by an AI agent has a Take over button. Only an organization admin can use it; for anyone else the button is disabled and says so. The button is absent on a chat that a workflow is driving, since there is no agent to take it from.

Taking over is one-way. The agent doesn't answer again in that chat, and there is no way to hand it back. Before anything else, the visitor sees the agent say I'm going to bring in a colleague who can help you from here. One moment. and then a notice reading You're now chatting with [name]. See what the visitor sees. The name follows your organization's chat name settings.

Once you hold the chat:

• A Write to the visitor… box appears at the foot of the tile. Enter sends and Shift+Enter starts a new line. Messages can be up to 4,000 characters.
• The visitor sees a typing indicator while you type.
• Anyone else watching the tile sees Handled by [name] where Take over was, and has no composer.
• Your messages and the notice appear in the transcript, each labelled with who said it.
• End Chat is in the bar at the foot of the tile. It asks how the chat ended, the same way ending a call does, then ends the chat and records your answer. Only the supervisor who holds the chat sees it.

When the person who took a chat over leaves

If the supervisor holding a chat is no longer watching it and the visitor goes unanswered for longer than your organization's threshold, the chat gets a Visitor waiting alert. The threshold is Flag a chat as abandoned after on the Organization page: 3 minutes unless an admin changes it. It counts from the visitor's oldest unanswered message, or from when the current holder took the chat, whichever is later. A holder who is still watching the chat never triggers it.

Another admin then sees Pick up in place of Handled by [name]. Picking up works like taking over: the visitor sees a new You're now chatting with [name] notice, with no second colleague line. Nothing hands the chat back to the agent.

After a supervisor takes a call over

Join needs your microphone, and asks for it before anything else. If your browser blocks it, there isn't one, or another app is using it, the call stays with the agent and a message says which. If your microphone stops during the call, a headset disconnecting say, the page switches to another microphone if there is one and tells you. If there isn't, a message says the caller can't hear you, with Try Again, and a microphone you connect is picked up by itself. Keep listening until Joining… finishes: if listening stops first and the take-over still goes through, the caller can't hear you, and a message says so.

When a supervisor takes a call over, the agent finishes the sentence it's on, for up to 4 seconds. With Allow barge-in on, a caller who talks over it stops it at once, as at any other time. The agent then waits up to 10 seconds for the caller to finish speaking, tells them a colleague is joining, and leaves. With Allow barge-in on, a caller who talks over that line hears it again once they finish, so it's the last thing they hear from the agent. What the caller says while the agent waits is in the transcript. Join reads Joining… until that's done: usually a few seconds, and longer while the caller is still talking, up to about 25 seconds.

After the agent leaves, the transcript carries on. A line reads [name] took over the call. The agent left., and what the caller and the supervisor say after that follows it, each line labelled with who said it. The caller and the supervisor are transcribed separately, so a line is never put down to the wrong person. A line appears once its speaker finishes it, so it can trail the audio by a few seconds. The column stays live until the call ends.

At close-out the supervisor can say why they stepped in. Why did you step in? offers Agent stuck or looping, Wrong or missing information, Caller asked for a person, Complaint or escalation, Compliance or risk, and Other. It is optional, and more than one can be picked.

Once the call is over:

• Caller memory takes in what was said after the take-over, so the caller's next call knows that a person stepped in and how the call ended. This happens only when the agent has caller memory switched on.
• Coaching reviews the agent once the call has ended and been closed out, without waiting for its usual count of calls, and reads why the supervisor stepped in and what they said. See Coaching.
• Call Logs shows who took the call over, the close-out, and the rest of the transcript. See Insights.

What was said after a take-over is kept as long as the call's own transcript: 90 days when the agent stores transcripts, whether for transcript retention, post-call analysis, or a wrap-up tool, and otherwise an hour, long enough to watch the call. Erasing a caller's data erases it along with their calls.

Reading the status values

Calls. Nearly every call shows Connecting for a fraction of a second and then Live. That's the agent starting up, and it isn't worth watching. A row that stays on Connecting is worth acting on: the caller is connected but hearing nothing. One such row is usually a single unlucky call. Several at once, across different agents, point at the platform rather than at any one agent's configuration — that's the point at which to raise it.

Chats. Active means a message landed in the last five minutes: someone is very likely still there. Idle means the conversation is still open and resumable, but nothing has been said for a while. Idle is the normal resting state of a chat, not a fault — most visitors read an answer and leave without closing anything. A board with a dozen idle chats and no active ones means your website was busy earlier, not that something is stuck.

Because idle chats accumulate through the day, the count beside the search box reports the two separately. The active number is the one that tells you how busy your agents are right now.

Reading the sentiment badge

A row can carry a second badge beside its status, saying how the person on the other end sounds. An open transcript shows the same badge in its header. There are three:

Most rows carry none of these, and that is the normal case. Nothing appears for an ordinary conversation, and nothing appears before there is anything to say — a call that has only just been answered, or a chat whose first message is still being answered, has no badge yet. A conversation that was already under way before this arrived may not get one at all.

Escalating is not the same as Negative, and it is the more useful of the two. A caller can be perfectly polite while asking three times to be put through to a person. That reads as neutral words and high escalation, and it is usually the conversation worth opening first. The two are measured separately for exactly that reason. When both apply, the row shows Escalating.

It reads the words, not the voice. There is no analysis of tone, volume, or how somebody sounds — a call is read from its transcript exactly as a chat is read from what was typed. So it inherits the transcript's mistakes: speech recognition is imperfect, particularly on names, numbers, and accents.

It keeps up with the conversation rather than waiting for the end of it, but it is not instantaneous. A chat is re-read once each message is answered, and can take a few seconds to reach the board after that. Read a badge as "a moment ago", not "right now".

The same badge appears on the caller's turns when you try an agent out from the agent editor, so you can see how a wording change lands before it reaches a real caller.

Treat it as a hint about where to look, not a verdict. It is there to tell you which of fifty live conversations to open first. It is not a measure of how a conversation was handled, and it is not something to hold an agent to.

When Party shows a dash

For a call, a dash means the call arrived with no caller id. That's normal for some carriers and for blocked numbers.

For a chat, a dash means the widget the conversation started in has since been deleted. The conversation itself is unaffected and the visitor can still continue it.

Note that chats identify the widget, never the visitor. A website visitor is anonymous, and this page deliberately publishes nothing that would identify one.

What this page is not

Coaching and taking over a chat are being rolled out. Where they aren't enabled for your organization, you can't coach the agent or message the visitor from here.

There is no history here. A call appears when it starts and disappears when it ends; a chat appears while it is going and goes when it ends or falls quiet. Nothing on this page is a record you can come back to later.

To look at calls that have already finished, with transcripts and post-call analysis, use the agent's Recent Calls (or Recent Chats), or Insights → Call Logs for workflow-level reporting.

"This list may be incomplete"

If a warning appears at the top of the page, part of the platform couldn't be reached when the list was last read, so some conversations may not be shown in the sidebar. The conversations themselves are unaffected. This is about what the page can see, not about what is working.

It normally clears on its own within a few seconds. If it persists, the count beside the list is an undercount, so don't use it to conclude that volume has dropped.

"Delayed Updates"

Nothing normally appears beside the conversation count; a working list shows nothing there. If a Delayed Updates badge appears there, the page has lost its live connection and fallen back to re-reading the list every few seconds. Everything still works; rows may just be a few seconds behind. It reconnects by itself.

An open transcript has its own version of this. Its badge normally reads Live; if it reads Delayed, that column's connection keeps dropping and messages may be missing from what you can see. The rest of the page is unaffected, and it reconnects by itself. A column showing Ended is not a fault — that call has finished.

Who can see it

Anyone in your organization. The Supervisor section is not restricted to administrators, because supervising conversations isn't an administrative task.

Note what that grants. Everyone who can open this page can see callers' phone numbers and read what callers and visitors say, including anything they volunteer: account details, addresses, health or financial information.

Reading it from the API

The same data is available programmatically: a single read of the list, a stream of changes to it, and a stream of one conversation's transcript. Each conversation carries a channel of voice or chat, which is what decides the rest of its fields, and — once there is a reading — a sentiment carrying the same labels the badge is drawn from. See listLiveConversations, streamLiveConversations, and streamLiveTranscript in the Voice Agent API.