Agent — overview

This is where you talk to the agent and watch it work: chats, topics, journals of background runs and the service screens. The menu item carries a badge of unread agent messages.

The left panel

Group	Items
Chats	All chats · Agent chats
Topics	your topics
Channels	channels, if any

At the bottom — New chat, with a chevron for more: new topic, new channel, new automation. The gear in the header opens the agent settings.

Chats

All chats are ordinary conversations grouped by time — today, yesterday, earlier — with search and an unread count. Row actions: open in a new tab, delete.

Agent chats are the ones the agent started itself, with a switch:

• Conversations — the agent needed to ask you something;
• Journals — records of background runs.

The run journal

A journal is not a conversation but a record: what the agent did, with which tools and how it ended. Replying there is pointless — there’s Discuss in chat, which opens a proper conversation with the run’s context.

While a background agent works or finishes, a status strip appears in the chat — “background agent working” / “background agent finished” — with Open and Hide. If it hits a question you’ll see “the background agent is waiting for your decision”.

Topics

A topic is an isolated stream of conversation: its own history, tone and role. The main topic cannot be deleted.

Creating one asks for a name (up to 100 characters), an icon (emoji) and a brief for the agent (up to 500 characters) — what this topic is about.

Topic settings:

• name and description — the description is for you, not the agent;
• tone — applies to every topic;
• agent role — who it is here: “you are a practising lawyer, cite the articles”;
• model override — a specific model for the topic; empty inherits the settings;
• tool permissions.

Note

In the main topic the role and the model override are unavailable — otherwise the agent loses its universal context. Name, description and tone are free to change.

Deleting a topic deletes all its conversations and cannot be undone.

Tool permissions

Category	What it covers
Reading data	searching and reading your objects
Creating and editing notes	writing to notes
Sending mail	outgoing email
Web search	searching the internet
Fetching pages	reading web pages
Running code and shell	executing commands
MCP tools	external tools
Finance operations	writing to finance

Values: Auto — no confirmation, Ask — confirm each time, Deny — block.

If a conversation drifts, the agent may offer to move it into its own topic — the Create button sets it up immediately.