Chatting with agents
The Inbox is where you and your agents talk. It is a two-way channel: agents surface work that needs your input, and you send instructions, answers, and guidance back. Think of it as the group chat for your digital workforce — one thread per topic, one agent per thread.
You reach it from the sidebar under Status → Inbox.
The Inbox is for conversation. When an agent needs a formal yes/no on a specific action, that lands in Approvals instead. The two are related but separate: chatting shapes how an agent works; approving clears a specific pending action.
Who can use it#
Chatting and giving guidance is an Operator-and-above capability. Owners, Admins, and Operators can open threads, reply, and hand over tasks. Viewers have read-only access — they can follow along but cannot send messages. See Team & roles for the full matrix.
The layout#
The Inbox has three regions:
| Region | What it shows |
|---|---|
| Agents (left rail) | Every agent you have a conversation with, plus an unread badge per agent |
| Conversations (middle) | The threads for the selected agent, newest activity first |
| Thread view | The open conversation — messages, the reply composer, and (for escalations) a context panel |
Selecting an agent filters the conversation list to that agent. Selecting a conversation opens the thread and drops your cursor straight into the composer so you can start typing.
Two kinds of conversation#
Threads come in two flavours, and they behave a little differently.
Direct conversations (chats)#
A chat is a conversation you start. Click New chat (top right of the conversation list) and the newest-selected agent gets a fresh thread titled with today's date. Say hello, ask a question, hand over a task — the agent replies right in the thread. The footer reads "Direct conversation — the agent replies right here."
Escalations#
An escalation is a conversation the agent starts because it hit something it wanted your input on. These carry extra context: a risk level, an impact summary, an estimated cost, and the affected systems, all shown in a panel on the right. The composer footer reads "Sent as guidance for this escalation — the agent knows the context", because anything you send is attached to that specific situation.
Escalations also show a "Confidence this will get a reply" meter in the header — the agent's read on how likely your input is to move things forward.
An escalation is the conversational sibling of an approval. An approval asks "may I run this exact action?"; an escalation asks "here's a situation — how do you want me to handle it?" You answer escalations with guidance, not a single button.
Reading a thread#
Agent messages are typed, so you can see the shape of the agent's thinking at a glance:
| Message type | Label | Meaning |
|---|---|---|
| Observation | Observation (eye icon) | Something the agent noticed |
| Reasoning | Why? (brain icon) | The agent's rationale — click to expand the full reasoning |
| Recommendation | Recommendation (lightbulb) | A suggested course of action, often with quick-action buttons |
| Action result | Done (green check) | Something the agent completed, with the outcome |
| Reply | (plain bubble) | A normal chat reply from the agent |
When an agent replies about an action it took in a connected tool, the message carries a small provider chip — the tool's logo, the channel or recipient, and the subject — so you can see it sent that Slack message to that channel, or that email to that person.
Your own messages appear on the right in a dark bubble. If you used Reply to answer a specific message, your bubble shows the quoted text above it.
Quick actions on every message#
Hover any agent message to reveal three actions:
- View — open the full message in a dialog (useful for long replies).
- Reply — quote this specific message; your next send is aimed at it and the agent is told exactly which message you answered.
- Copy — copy the message text to your clipboard.
Sending a message#
- Open a conversation.
- Type in the composer at the bottom. Press Enter to send; press Shift + Enter for a new line (the box grows as you type).
- To answer a specific message, hover it and click Reply first — a blue "Replying to…" banner appears above the composer with the quoted text. Press Escape or the ✕ to cancel the quote.
While the agent works on its response, you'll see a typing indicator: "[Agent] is working on a reply…".
Quick-action buttons#
Recommendation messages often end with buttons like Acknowledge, Proceed, or other suggested next steps. Clicking one sends it as a control instruction — a fast way to steer the agent without typing. The first button is the primary (recommended) choice.
Live message notifications#
You don't have to sit in the Inbox to catch new messages. When an agent sends one, a toast slides in at the bottom-right of the screen anywhere in the dashboard. Click it to jump straight to that thread.
For alerts even when the dashboard tab isn't focused, turn on Desktop notifications — see Notifications. Which events notify you (agent updates, approval requests, errors, daily summary) is configurable per channel.
Read state and unread counts#
- Opening a conversation marks it read, and the read receipt syncs across your devices and the sidebar badge.
- Each agent in the left rail shows a count of its unread conversations; each conversation shows a numeric badge of unread messages.
- New activity after you've read a thread re-arms the unread dot.
Managing conversations#
Every thread has a ⋮ menu (in the conversation list on hover, and in the open thread header):
- Rename — give the conversation a clearer title.
- Archive — move it out of the active list. Archived threads aren't deleted.
To bring one back, click Archived at the top of the conversation list, find the thread, and click Restore.
Older history#
Long threads load the most recent messages first. Scroll to the top and click Load earlier messages to page back through the history; your reading position is preserved so nothing jumps.
What's next#
- Approvals — clear the specific actions your agents are waiting on.
- Notifications — control message toasts and desktop alerts.
- The agent workspace — see everything one agent is doing, tab by tab.