Conversations
Replay end-user sessions across every channel your agent serves
Where Monitoring is for engineers debugging individual runs, Conversations is for product, support, and ops teams reviewing how end-users actually experience your agent. Every multi-turn session — Slack, Teams, Telegram, WhatsApp, web, API — is captured here as a thread you can replay.
What's a conversation?
A conversation is a sequence of runs that share session state for a single end-user. The agent's working memory, retrieved knowledge, prior tool calls, and turn-by-turn context are all preserved across the thread.
Conversations are channel-aware: a Slack thread is one conversation, a Telegram chat is another, a logged-in user on your web app is a third. The user's identity (when known) is attached.
The conversation list
On each agent, the Conversations tab shows every active and historical session, with:
| Column | Meaning |
|---|---|
| Channel | Slack, Teams, Telegram, WhatsApp, web, API, etc. |
| Name | The conversation's identifier (e.g., the channel/thread name or chat title) |
| Messages | How many messages so far |
| Last message | When the most recent message arrived |
Filters
- Channel — isolate one delivery surface
- Date range — narrow to a specific window
- Search — find a conversation by name
The conversation viewer
Open a conversation to see the full thread, message by message:
- User messages with attachments preserved (images, files, links).
- Agent responses rendered the way the end-user saw them.
- Tool calls the agent made on each turn, expandable inline.
- Per-turn trace — click any agent message to jump to the underlying trace with all the engineering detail.
- Knowledge citations when the agent retrieved from your knowledge base.
Actions
- Add to evaluation dataset — turn a conversation into a regression test in one click (Evaluations).
Privacy & retention
- Conversations are scoped to your organization and only accessible to teammates with appropriate roles (Team & Roles).
- Retention follows your plan; once retention elapses, conversation data ages out.
- For PII / compliance use cases, see Security.
Patterns
- Daily sample. Skim a handful of recent conversations. You'll catch tone drift, formatting bugs, and silent quality issues that aggregate metrics miss.
- Promote to dataset. Any conversation that surprises you (good or bad) belongs in your evaluation dataset. Add it.
- Watch the long ones. Conversations with many turns often signal a confused agent. Look for repeated tool calls or a memory not being updated.
Next steps
- Monitoring — engineering-grade traces per run
- Connectors — the channels conversations come from
- Memory — what the agent remembers across turns
- Sessions — context-window management for long threads