AI Agent

Agent Conversation

Pro+Agent Available on the Pro+Agent tier.

The conversation panel is how you talk to your agent. It accumulates your messages, the agent’s responses, tool activity, and system notices in a single thread that persists for the duration of the agent session.

Open and close the panel

Open the agent conversation panel using the toolbar button. Close it with the X button in the panel header, by pressing Escape, or by clicking the backdrop outside the panel.

Closing the panel does not stop the agent — it keeps running in the background. Reopen the panel to continue where you left off.

Send a message

Click the message input at the bottom of the panel and type your message. Press Enter to send. Use Shift+Enter to insert a line break within a single message without submitting it.

The send button is disabled when the input is empty or contains only whitespace.

After you send, your message appears right-aligned in the thread. The agent’s response appears left-aligned once it begins streaming.

Read streaming responses

The agent streams its response progressively. A blinking cursor inside the assistant bubble indicates the response is still generating. Each chunk of text appends to the existing bubble — not as a separate message.

While the agent is generating:

The message input shows “Agent is thinking…” and is disabled
The send button becomes a Stop button (square icon)
The status label reads “Thinking…”

When the response finishes, the blinking cursor disappears, the input re-enables, and the status returns to “Ready.”

Tool call indicators

When the agent uses a workspace tool, a badge appears below the thread entries showing the tool name and a “running…” label. When the tool completes, the label changes to “done” and the badge clears after two seconds. If a tool fails, the label shows “failed” in red before clearing.

The status label also reflects tool activity — while a tool runs, it shows “Using search_pages…” (or whichever tool is active).

Permission request dialogs

Some agent actions require your explicit approval. When the agent needs a capability it hasn’t been granted, a permission dialog appears inline in the thread. The dialog shows the tool name, the required capability, and the agent’s explanation.

Three options are available:

Option	Effect
Allow Once	Grants permission for this operation only. Future requests for the same capability will prompt again.
Always Allow	Grants permission for the rest of the current session without further interruptions.
Deny	Blocks the operation. The agent receives the denial and responds accordingly.

If the panel is closed when a permission request arrives, a banner appears at the bottom of the screen. Click View in the banner to open the panel and respond. Dismissing the banner hides it without denying the request — the dialog remains available when you reopen the panel.

System messages

Centered messages in the thread mark significant events:

“Session interrupted.” — you clicked Stop while the agent was generating.
“Error: LLM provider not configured.” — the agent ran but no API key was set up.

System messages are informational. You can send a new message immediately after one appears.

Clear the conversation

Click the trash icon in the panel header to clear all messages. The empty state reappears. Clearing the thread is local — the agent session continues running. The agent’s persistent memory is unaffected.

Conversation lifecycle

Behind the UI, each agent session is tracked as a conversation with a defined lifecycle:

Status	When it applies
Active	The agent session is running.
Idle	The session ended normally. The conversation is available for review in memory.
Archived	You archived the conversation. Excluded from default views.

Conversations belong to a channel — a topic-based grouping within your workspace. The default “general” channel is used unless a specific channel is active.

Keyboard reference

Action	Key
Send message	Enter
Line break in message	Shift+Enter
Close panel	Escape