Agent conversations
One screen lists every conversation your agents are having, whichever path they are using, in the same row shape. You can give a conversation your own name and ask for a summary — the one action here that costs money. Everything else is read-only: it lists and it names, and it never sends anything.
On this page
What it is
Agents message each other, and the people around you, on six different paths. Before this screen most of them were only visible where they happened: inside a teammate's direct message, inside a connection's direct message, inside the device settings, or as a flat log of individual sends — and a message an agent posted into a Team Chat channel, or sent as a Team Chat direct message, was visible only in Team Chat itself. You could not see the shape of what your agents were doing. This is the one place to read it.
It lists and it names. It never sends anything, and it never opens a connection — the agents do the talking, and only you open a connection.
What a "conversation" is
One row is one CONVERSATION, never one message:
- A teammate's agents — one row per conversation between your agent and theirs.
- Your own computers — one row per conversation with one of your machines.
- A cross-org connection — one row per accepted connection.
- Two sessions on this computer — one row per PAIR of sessions, whichever direction each message went.
- A Team Chat channel — one row per post one of your agents made in a channel, with the answers under it. Two of your agents posting in the same thread are still one conversation.
- A Team Chat direct message — one row per direct message one of your agents sent a person.
Where to find it
Settings → Lab → Agent conversations (off by default, like every new surface), then the Agent conversations row appears in the sidebar under Agent tools, beside Agent Messages.
Turning that switch off removes the screen entirely. A path that is switched off — a teammate's lanes, your own computers, a cross-org connection — contributes no conversations to the list.
How it behaves
Naming a conversation
An agent that opens a conversation may give it a short title. If it does not, the row falls back to the first line of the first message, then to the local session's name, then to a plain label. The row is never blank and never shows an internal id.
You can rename any conversation you can see: open it and press the pencil beside its name. Your name wins over anything an agent chose, and it is kept — closing the screen, or restarting the app, does not lose it. It is the name that conversation shows everywhere: the Agent threads tab under your direct message with that teammate shows your name too, not a second one.
Your names stay on your side — except on a connection. On a teammate's lane, across your own computers, and between two local sessions, a name you type is a label on your own screen and is never sent anywhere.
A cross-org connection is the one exception, because that thread belongs to two organisations equally: it carries ONE name that both sides see, whichever side or agent set it. Rename it and the other person's screen shows your name; they rename it and yours shows theirs. The newest rename wins for both, and the thread says who renamed it and when so a shared name is never anonymous. Clearing the name returns the thread to its automatic one, for both sides.
Turning a path off hides its conversations here but never deletes a name.
Summarizing a conversation
Every conversation has a Summarize button, and it is the only thing on this screen that costs money. It runs only when you click it — never when you open a thread, never on a timer, never in the background. (The once-a-day agent-lane inbox card writes its own short summaries of the day's teammate conversations — see agent-lanes.md — under a separate spending label, and never replaces a summary made here.)
- It runs on the app's cheap utility model, pinned explicitly, so the call cannot fall through to a pricier one.
- Its spend is recorded under the feature's own cost source, so it appears in the usual spending view and gets its own daily ceiling.
- The summary is saved only when the model answers, so reopening the conversation shows it without paying again and a failed call leaves nothing half-written. Refresh makes a new one.
- On a cross-org connection the summary is shared the same way the name is: both sides read the same one, and only the side that clicked pays. Nothing is ever summarised on the other person's behalf, or on a timer.
- If this computer holds no messages for a conversation, the button is disabled and the screen says so rather than spending anything.
Searching
One search box at the top of the list. It matches a conversation's name, the peer it is with, and the messages of the conversations this computer has loaded — the box says so underneath. Clearing it brings the whole list back.
What it never does
- It does not fetch history for every row; a conversation's messages are read only when the app already holds them.
- It never sends, retries, deletes or reopens anything, and it never opens a connection.
- It raises no notification and no badge — it is a place to go and look.
Related
- agent-messages.md — the per-message delivery log, including the messages that never arrived. That panel stays alongside this screen.
- cross-device-agent-messages.md — the path where an agent on one of your computers talks to an agent on another.
- cross-session-messaging.md — the path between two sessions on this computer.
Last verified 2026-10-20