Chatting with an agent

Chatting with an agent is the single most-used surface in Claos after the inbox. You open the agent, ask, and watch the work happen. This article covers what to attach, what to expect as the response streams in, and where the conversation lives once it's over.

Opening a chat

Click an agent on the Agents page. You'll land on the agent's detail page; the chat sits in the main area, ready for your first message. If a link sent you here with a starter prompt attached (some tour and onboarding links do this), the message is pre-filled — you only need to hit send.

What you can attach

Three things you can bring with a message:

Files. Drag a file into the input, or use the attach button. Supported types are PDF, CSV, DOCX, TXT, and MD. Files are uploaded and read by the agent for that turn — they're not stored as workspace pages.
Mentions. Type @ to mention a teammate from a list. Useful when you want the agent to address a specific person in its output ("draft a reply to @sam") or to give the agent a hand-off.
Implicit context. If you opened the chat from a project, a page, or a task, that scope is already attached. The agent knows what you're talking about; you don't have to recap.

Suggestion chips

Below the input you'll usually see four to six starter prompts tailored to the agent. They come from the agent's saved suggestions or, if none are set, from a generator that reads the agent's description and tools. Click one to send it as your first message — handy when you're not sure how to phrase the ask.

Reading a streamed response

Responses stream as the agent works. You'll see them in this order:

Thinking — a brief italic line describing what the agent is doing right now.
Tool calls — for example, "searching your workspace…" — each with a status badge while it runs.
Artifacts — pages and tasks the agent is creating or editing, surfaced as previews in a side panel as soon as they're produced.
The reply — the agent's text, typed out character by character.

You can stop the stream at any time with the Stop button next to the input. The partial output stays in the conversation; nothing is discarded.

Where past chats live

Every chat with an agent is saved in the conversation sidebar on the left of the detail page. The sidebar lists both your chats and the agent's past runs, with filter pills above the list — All / Chats / Runs — to narrow. Conversations are grouped by Today, Yesterday, Previous 7 days, and Previous 30 days, and there's a search box for finding a specific thread. Each conversation has its own URL (the ?cid=… parameter on the page), so you can copy a link straight to a single thread.

What a chat can produce

What ends up in your workspace depends on the agent's instructions and which tools it can use. The common outcomes:

A reply in the chat. Text only, nothing saved beyond the conversation. Most one-off questions end here.
A page in the knowledge base. When the agent creates or edits a page, the page opens in the side preview while it's being written and is saved to a folder — either the agent's default output folder (set in its settings) or one it picks at runtime. Pages can carry rich content: headings, lists, tables, callouts, and code blocks, all rendered as part of the page rather than as separate artifacts.
A task in a project. When the agent creates or edits tasks, each one appears as a clickable chip; clicking opens the task in the same side preview. Useful for asks like "turn these meeting notes into tasks" or "break this brief into a workplan."
A comment on a task. When the agent is working against a specific task — for example, a task you assigned to it — it can post conclusions and tool-call summaries back to that task's comment thread.

Tips and gotchas

The chat surface here is per-agent. There's also a workspace-wide AI chat — the Chat tab in the sidebar — which is a separate surface for talking across all your agents. Don't confuse the two.
You can also chat with an agent from Slack, Telegram, Discord, or WhatsApp once you've linked a messaging channel — same conversation shape, different surface. See What are messaging channels?.
Sent messages can't be edited, and there's no regenerate button on a previous reply. If a response isn't quite right, send a follow-up — agents iterate readily, and they'll keep the earlier turn in mind.
File attachments are read for the turn you attach them to. If you want the agent to know a document long-term, save it as a knowledge base page and add it as a skill instead.
If a chat is scoped to a project (because you opened it from one), the agent treats that project as the default context for its tool calls. To switch context, use the project selector at the top of the chat.