Chat
Chat is where you talk to your agents and teams directly. Every conversation is a chat session: a saved, resumable transcript bound to one agent or one team, in one project. Agents can call tools, ask for your approval, delegate to other agents and hand the conversation on, all inside the session.
Use chat to try out an agent before you put it in a team or workflow, to get work done with an agent day to day, and to answer approvals that a running agent is waiting on.
Where chat lives
The same conversation engine appears in several places. Each one shows a session in a different layout.
| Surface | How you reach it | Best for |
|---|---|---|
| Chat page | Chat in the left nav, or /chat | Full-width conversations, your conversation list, search, and the live run panel |
| Chat drawer | The chat icon in the top-right header | Talking to an agent while you work on another page. See Chat drawer |
| Pop-out chat | Pop out chat in the drawer header | A floating, movable chat window over the page |
| Builder chats | The detail pane of a library page (for example Agents, Teams or Functions) when nothing is selected, and "build this in chat" buttons such as Talk to the … or Ask about … (for example on Workflows) | Asking the agent that builds that kind of thing to create or change one for you |
| Program assistant | The program wizard | Drafting a program with an assistant |
| Microsoft Teams | A Teams bot you register | Letting people chat with an agent from Teams. See Microsoft Teams bots |
The Chat page and the drawer share your conversation list. A conversation you start in the drawer appears on the Chat page, and Open full view in the drawer moves the open conversation to the Chat page without restarting it.
Who you can chat with
You start a chat with exactly one of:
- An agent. One agent answers. You can override the model for this chat (see Model per chat).
- A team. The team's orchestration pattern decides which agent answers each turn. See Teams.
The agent or team is fixed when the session starts. To talk to someone else about the same conversation, hand it off.
When you open a new chat, FYAI preselects the agent you used most recently. If you have none, the picker asks you to Select an agent or team. The picker lists PINNED entries first, then RECENT, then all TEAMS and AGENTS alphabetically. Star an entry (Pin to top) to keep it at the top; pins are saved to your account.
Permissions
| Chat with | Who can start it |
|---|---|
| A team | Project Editor or above, or an organisation Owner/Admin |
| An agent | Project Editor or above, or a member who has been granted run access to that agent |
Run access is granted per member under Settings → Projects using Manage run access. A run grant on an agent lets the member start a one-to-one chat with that agent without broader edit rights. See Members and roles.
If your organisation's subscription has lapsed to read-only, the message box shows Read-only — upgrade to continue and you cannot send.
Privacy and visibility
- Chats are private to the person who started them. Your conversation list, search results and "needs you" counts only ever include your own chats, even for administrators.
- Organisation Owners and Admins can open another member's chat if they have its link (for oversight), but it never appears in their own list.
- Chats are per project. Switching project shows that project's conversations.
- Copy link gives you a
/chat/…URL. It only opens for you (and for organisation admins).
Chats and runs
Agents also run without a person typing: workflow nodes, program activities, webhooks, schedules, sub-agent delegations and Claude Code delegations all create sessions. FYAI calls these runs and keeps them out of your conversation list. They appear on the Activity page instead (see Managing conversations).
Sessions created by a hand-off are treated as chats, so they stay in your list.
Session lifecycle
A session is created when you send your first message. Its transcript is saved as it goes, so you can close the browser and come back.
Each conversation in the list carries a coloured dot:
| Dot | Meaning |
|---|---|
| Waiting on you (amber) | An approval in this chat is waiting for your answer |
| Running now (pulsing green) | The agent is working right now |
| Active (green) | Used in the last 24 hours |
| Failed (red) | The last run failed recently |
| Dormant (grey) | No activity for more than 24 hours |
A bold title means the chat has replies you haven't read.
Runs survive a dropped connection. If you leave a chat while the agent is working and come back, a banner says either This chat is still running on the server. or This chat was left while the agent was working. Select Reopen to reattach and see how it finished.
Titles
A new chat is titled from the first few words of your first message. Once the agent gives its first real reply, FYAI asks the chat's own model for a short title (at most six words) and replaces it. Tool output and the agent's greeting are never used for the title. If you rename a chat yourself, the automatic title never overwrites your name.
Model per chat
For a one-to-one agent chat you can pick a different model from the dropdown next to the agent picker before you send the first message. The first option, Default (…), is the agent's own model. The list shows the models configured in your project.
Once the session exists the model is locked (Model is locked for this chat — start a new chat to change it). Team chats have no model override; each agent uses its own.
If the model provider starts rate limiting, the chat shows The model provider is rate limiting requests. For an agent chat you can choose Switch model, which stops the current run and starts a new chat on the model you pick. The old transcript stays in your history, and no tool calls are replayed.
Focus
When you open chat from an entity (for example Ask about … on a workflow), the chat carries a focus chip such as workflow: Weekly report. The agent is told about that entity and will act on it unless you say otherwise. Clear the chip with its × to remove the focus.
Greetings and opening suggestions
What a new chat shows before you type comes from the agent or team:
| Setting | Where | Effect |
|---|---|---|
| Greeting | Agent form | The agent's opening message. If empty, the chat starts with Welcome to the chat! |
| Opening suggestions | Agent form, one per line | One-tap prompts shown until you send your first message |
| Introduction Message | Team form | The team's opening message |
On the Chat page, a new chat also shows a welcome panel: a greeting, What should we work on?, the agent's opening suggestions as cards (plus Run my first automation), and an or talk to row of other agents and teams (pinned and recent first, +N more for the rest). Close it with Hide this; it stays hidden for your account on every device.
Memory in chat
If an agent has memory enabled, FYAI may give it relevant memories from earlier conversations. Memories used (in the chat's More actions menu, shown when there are any) lists the memories injected into the current session, where each came from (From this session or From session #…), and lets you Forget this memory so the agent is no longer given it. See Memory.
Session stats
Session stats (in More actions) shows what the current session has cost and done once it has completed a turn:
| Stat | Meaning |
|---|---|
| Turns | Conversation turns |
| LLM calls / Tool calls | Model requests and tool executions |
| Est. cost | Estimated model cost |
| Input tokens / Cached tokens / Output tokens | Token usage |
| Wall time / Model time / Tool time | Where the time went |
| Duplicate calls, Polling calls, Longest failure streak | Signs an agent is looping or struggling |
Tool failures by class breaks down failed tool calls. For usage across all sessions, see Usage and billing and Observability.
Chat-related settings
| Setting | Where | Who | What it does |
|---|---|---|---|
| Show Debug Info | Chat More actions → Settings | You | Shows low-level connection and protocol messages in the transcript. Off by default |
| Show Function Info | Chat More actions → Settings | You | Shows tool-call steps in the transcript. On by default. Delegation and artifact cards always show |
| Pin … to top | Chat More actions → Settings, or the star in the picker | You | Keeps an agent or team at the top of the picker |
| Clear Chat History | Settings → Data & Privacy | You | Permanently deletes all your own conversations in this project. Chats that are still running are skipped. Cannot be undone |
| Manage run access | Settings → Projects, member list | Project admins | Lets a member chat with specific agents |
| Teams Bots | Settings → Teams Bots | Organisation | Connects an agent or team to Microsoft Teams. See Microsoft Teams bots |
| Greeting, Opening suggestions | Agent form | Builders | What a new chat with the agent shows first |
| Experience Agents | Settings → Experience Agents | Platform administrators | Which agent or team answers each "build this in chat" entry point |
Limits
| Limit | Value |
|---|---|
| Attachments | Images only, pasted or dropped, up to 5 MB each. The model must support image input |
| Earlier messages on the Chat page | Loaded 50 at a time with Load earlier |
| Undo after deleting or archiving a chat | About 10 seconds |
| Search results | 20 per search |
| Agent hand-offs started by an agent | One per conversation |
Chat from the API
Everything here is also available programmatically: starting and ending sessions, sending messages, handing off, searching history and deleting history. One API-only feature is session justification — recording a business reason against a session (PATCH /chat/{id}/justify, or the justify_chat_session MCP tool) — which appears in the Assurance Pack evidence. See the Platform API and Platform MCP server.
Related
- Chatting with agents — start a chat, send, stop and steer a run
- Approvals and hand-offs in chat
- Managing conversations — rename, pin, archive, delete, search
- Chat drawer
- Microsoft Teams bots