Sessions
Every chat the agent had, by channel, with its transcript and token counts.
Sessions lists the agent's chats, newest first. Open the agent and go to Agent → Sessions. Pick a chat to read it.
Channels
Filter by channel, or pick All channels.
| Channel | Where the chat came from |
|---|---|
| Team chat | Open chat on the agent's card |
| API | POST /v1/chats with an API key |
| Widget | A visitor on your website |
| Webhook | A webhook call |
| Schedule | A schedule run |
| Eval | An eval trial, a judge, or an optimization run |
Each row shows the first message, the channel, the number of turns, and when it was last updated.
What is kept
A transcript shows each message, the tools the agent used, and errors. Its header shows the channel, when the chat started, and the tokens in and out.
Chats are kept in the agent's PostgreSQL and survive restarts. A chat not updated for 90 days is removed. Deleting the agent deletes its chat history.
An agent from before chat history was stored keeps chats in memory, and the tab says so. They reset when it restarts.
Add to eval
Add to eval turns a real chat into a test. Its first message becomes a test case's request. Pick A new eval or an existing eval, then save it. The reply that chat got is shown next to the test case while you write its checks. See Evals.
Limits of a chat
| Limit | |
|---|---|
| Message | Up to 8,000 characters |
| Turns | Up to 30 per chat |
| Reply time | A reply that takes longer than 120 seconds is stopped |
| Chats at once | 5 live chats per agent. A new one replaces the least recently used idle chat. If all are mid-reply, the API returns 503 with Retry-After. |
A chat with no messages for 10 minutes is put to sleep. The next message wakes it with its history.
API
GET /api/ai-agents/{pipeline}/{phase}/{name}/sessions lists chats (?limit, ?before for the next page, ?channel). GET …/sessions/{chat} returns one transcript. See API.