Messages between agents
When an agent prompts another session, either on the same relay or on a paired relay, the message lands in that session as a prompt like any other. This page shows how you tell it from your own messages, how the exchange is folded so it does not crowd the transcript, and what happens with the reply.
How an agent sends such a prompt is described under agent sessions (sessions on this relay) and relay_tool (the tool, and sessions on paired relays).
Where a prompt came from #
A prompt that an agent sent carries a badge above its text, so you can tell it from a message of your own.
| Sent by | The badge reads |
|---|---|
| An agent on a paired relay | ↗ from relay- |
| An agent on this relay | via agent · “session title” · model |
The session title is the one of the conversation the agent runs in. Parts the relay does not know are left out. Hover the badge for a sentence that says who sent it.
- From a paired relay, the badge is a link to the sending session on that relay, when the relay's address is known. It opens in a new tab.
- From this relay, the badge opens the conversation the agent runs in, as long as that conversation is still in your list. It shows that conversation's current title. If it is gone, the badge stays as plain text.
- In a shared read-only view, the badge is plain text and does not link.
Sessions an agent started #
A session that an agent created carries a marker too:
- From this relay: via agent · “title” in the conversation list and under the title in the header. Clicking it opens the conversation of the agent that started it.
- From a paired relay: ↗ relay-name in the conversation list. The header shows no marker for these.
The folded note Added in 0.9.10 #
An agent's prompt is folded into one row, so a session that other agents talk to a lot stays readable:
note from build-box · report builder · answered below
The row names the sender, the session it came from, how the turn stands and
the time of the prompt. For an agent on this relay the sender is the name of
this relay (this relay if it has none); the session is the current title of
the conversation the agent runs in. On a phone the row looks the same.
Added in 0.9.11 Only the note itself folds. The session's reply to it always shows, like a reply to you, because it may hold a line for you; so do the running turn and every card of that turn. In 0.9.10 the reply was folded into the row as well.
| Status | Meaning |
|---|---|
working… |
The turn is still running. |
answered below |
The session answered; the reply follows the row. |
failed |
The turn ended with an error. |
stopped |
The turn was stopped. |
no reply yet |
No turn is running for the prompt and none has answered, for example while it is queued or after it was cancelled. |
Click the row to open the note, with its badge, as it would look unfolded. Click again to fold it. Only the transcript is shown differently; nothing is deleted. A jump to the note, from search, a pin or an image's parent, opens the row.
The ⌄ button counts the note and its reply as two rows.
While the turn works #
A note that is still being worked on is open, so you can read along. When the turn is answered, failed or stopped, the row folds by itself. If you opened or folded the row yourself, it stays as you left it, also when the transcript is drawn again or you leave the conversation and come back. A reload of the page resets this.
What is never folded #
Nothing addressed to you is folded away: your own messages, the replies to them and the replies to agents' notes always show. An agent's prompt that was steered into a running turn is not folded either; it shows as part of that turn. A turn you steered one of your messages into folds its note like any other, since nothing of yours is behind the row.
How the reply goes back #
You do not reply to the agent. The session works on the prompt like on any
other, and its reply is handed back to the agent that asked, which waits for
it with send or wait, or reads it later. The session's agent is told so,
and is told where the prompt came from, so it ends with the result the other
agent needs.
The receiving agent also sees a line at the start of the prompt that says another agent is talking to it:
[Remote prompt from an agent on relay "build-box" · session "nightly digest" · claude-sonnet-5 · acting for the user]
The line is part of what the agent reads. The conversation does not show it on the message, because the badge carries the same facts.
Questions #
If the session asks a question while it works on an agent's prompt, the
question card waits in that session, like any other. You can answer it there.
The agent that asked also learns about the question, and a paired agent with
the Full level, or an agent on this relay, can answer it with
answer_.
Agents are told to ask you first, unless you told them to decide for
themselves. An agent cannot answer a question of its own session. The approval card for a
call that a session's agent makes to another relay is always yours to answer.
Notifications #
A turn that an agent started sends no "reply ready" and no "failed" notification: the agent is the one waiting for the reply, and it reports to you itself. Question notifications are unchanged, so a question that waits for an answer still reaches you.
Switching it off #
Accept prompts from other relays' agents under Settings → Relays switches off the incoming prompts from paired relays. It does not apply to agents on this relay; those are governed by the switch described under agent sessions.
By provider #
Nothing here depends on the provider of the session that receives the prompt. The badge, the note row and the notifications work the same for every provider.