Chatting with agents
This guide covers a conversation from the first message to the last: starting it, sending, steering the agent while it works, and reading tool calls, delegations and results. For approvals and hand-offs see Approvals and hand-offs in chat; for renaming, archiving and searching see Managing conversations.
Start a chat
- Select Chat in the left nav. (Or open the chat drawer with the chat icon in the header — see Chat drawer.)
- Select New chat at the top of the conversation list if a previous conversation is open.
- Choose who to talk to in the picker at the top (Select an agent or team). You can also pick from the or talk to row on the welcome panel. Type to filter; use the star to pin an agent or team to the top.
- Optional, agent chats only: choose a model from the dropdown beside the picker. Default (…) uses the agent's own model. You can't change it after the first message.
- Type your message, or select one of the opening suggestions, and press Enter.
The chat is saved as soon as the first message is sent, and the header switches from the picker to the chat's title.
Show me around (offered by the Platform Concierge) starts the guided platform tour instead of sending a message. Run my first automation launches a starter automation and takes you to its run.
Write and send messages
| Action | How |
|---|---|
| Send | Enter, or the round arrow button |
| New line | Shift+Enter |
| Dictate | The microphone button (Start listening / Stop listening). Uses your browser's speech recognition; available in browsers that support it |
| Attach an image | Paste it into the message box, or drag it onto the chat (Drop image or text). Images only, up to 5 MB each; larger files are skipped. Remove one with its × before sending |
| Drop text | Drag text onto the chat to insert it into the message box |
Drafts are kept per conversation while you switch between chats in the same browser tab.
If the agent's model can't read images, sending an image fails with This model does not support image input. Remove the image, or choose a vision-capable model for the agent.
Steer the agent while it works
While the agent is working, the status line under the transcript shows what it's doing (for example Thinking…, an agent name with what it's doing, or a tool name) and a token count. The Send button turns into a red Stop button.
Add to the conversation mid-run. Keep typing and press Enter. Your message is queued and the agent picks it up after its current step (a tool call already in progress finishes first). Use this to correct course without stopping.
Stop generating (composer button). The red button in the message box stops the stream in your window so you can type a correction straight away. It doesn't end the session; your next message goes to the same conversation.
Stop the run (header button). The Stop button in the chat header appears only while a run is in flight and ends it: running tools are cancelled and the session is closed. If a tool is still running you're asked to confirm (Stop this run?). After this the message box is disabled; start a New chat, or reopen the conversation from the list to continue it.
Cancel one tool. Expand a running step and select the × on the tool card (Cancel …). The agent carries on without that result. This isn't offered for Magentic teams, whose orchestrator copes badly with a cancelled tool; use Stop instead.
Read what the agent did
| You see | What it is |
|---|---|
| Agent messages | Markdown replies. Long replies collapse with Show more. Code blocks have Copy code |
| N steps chip | The tool calls behind a reply, grouped. Shows · running while in progress and · interrupted if the run ended mid-call. Select it to expand each call's Arguments and result |
| Artifact card | A platform object the agent created or updated (dashboard, agent, team, workflow, program, function, model config), with Open to go straight to it and Show details for the raw result |
| Delegation card | Agent delegated to specialist, the task, live progress, Working… with elapsed time, then Done or Failed. On the Chat page, View transcript opens the specialist's own conversation in a side panel |
| Approval card | Approval Required — see Approvals and hand-offs |
| Connect card | Connect … a service the agent needs — see Approvals and hand-offs |
| Progress ledger | For Magentic teams, a status card of the orchestrator's plan and progress |
| Execution environment banner | Execution environment "…" is … — LEA tools may be unavailable. The agent's execution environment isn't running, so tools that run there may fail |
To hide tool steps, open More actions → Settings and clear Show Function Info. Delegation and artifact cards always stay visible. Show Debug Info adds low-level connection messages, which is only useful for troubleshooting.
Double-click any message to open it in a larger popout. Select an image in the transcript to zoom it.
The run panel (Chat page)
On the Chat page on a desktop screen, a Run panel on the right shows the live run: Live with elapsed time, a Needs you list of pending approvals with Approve and Reject, the Agents involved (lead and specialists, working or done — select one to open its transcript), recent Steps, and the token count. Collapse it with Collapse and bring it back with Show run details.
More actions → Delegation map shows every agent the conversation delegated to as an org chart with live status. It is available in the drawer as well.
Edit a message
Hover over one of your own messages and select the pencil (Edit message). Change the text and select Send.
Editing rewinds the conversation: FYAI ends the current session and starts a new one with the messages before the edited one plus your new text. The original session stays in your list.
Rate replies
Hover over an agent reply and select Thumbs up or Thumbs down. Select it again to remove your vote. Ratings are stored against the message and feed agent quality signals.
Start over
- New chat (conversation list on the Chat page, or the + in the drawer) leaves the current conversation as it is and opens a blank one with the same agent (or, after a team chat, your most recently used agent) preselected.
- More actions → Clear chat asks Are you sure you want to clear the chat history?, ends the current session and gives you a blank conversation. The ended conversation stays in your list.
Connection status
| Status | Meaning | What to do |
|---|---|---|
| Connecting… | Opening the live connection | Wait |
| Reconnecting… Agents are still working in the background. | The connection dropped; the run carries on | Wait; the transcript catches up |
| Connection lost. Messages may not be delivered. | Reconnecting failed | Select Reconnect |
| This chat is open in another window. | The same chat took over in another tab or window | Select Use here to bring it back to this window |
A conversation can be live in only one place at a time. On the Chat page, the header chat icon focuses the message box instead of opening the drawer on top of the same conversation.
Related
- Chat — how sessions, privacy and settings work
- Approvals and hand-offs in chat
- Managing conversations
- Testing agents and teams — repeatable tests rather than ad-hoc chats