Observability
Conversations: read threads and score replies
Every conversation with your agents and workflows in one list, filtered by Live or Test traffic, thumbs down or low scores, with actions on each reply: Score, Run evaluator, Add to dataset, Send to review and Open trace.
Guide3 min readUpdated 31 October 2026
Conversations, under Observe in the left menu, lists every conversation with your agents and workflows, newest first, with its thumbs and scores. Open one to read the whole thread and act on any reply. It follows the agent picked in the top bar (a link with ?agent= wins); picking an agent leaves out workflow conversations.
The list
Each row shows when it started, the agent or workflow, the channel with a Live or Test pill, the turns (messages from the person), the first message, its feedback (thumbs up and down, and up to two score chips) and the last activity. 25 a page.
| Filter | Choices |
|---|---|
| Environment | Live (the default): the widget, chat links and the API, including workflow steps that call agents. Test: the Playground and agents' test chats. Or All |
| Channel | All channels, or Widget, Chat link, API or Playground |
| Period | Last 24 hours, Last 7 days (the default), Last 30 days or Last 90 days |
| Thumbs down | Only conversations with a thumbs down in the period |
| Score below | A score key and an average below a number from 0 to 100 (50 by default) |
| Search messages | Words in any message of conversations that started in the period |
Filters stay in the page's address, so a filtered list can be bookmarked or shared. A total ending in + means the filters looked at the newest 5,000 conversations in the period. Experiments never create conversations.
A conversation
- Messages: the whole thread (the newest 500 messages), each reply with its time, tokens, seconds and score chips.
- Details: the agent or workflow, the channel, when it started, the last message, turns, tokens, the conversation ID (with a copy button) and links to its traces.
- Feedback: every thumb and score, newest first, with who gave it (a person, an evaluator, a user, your API or thumbs), the comment and any correction.
Actions on a reply
| Action | What it does |
|---|---|
| Score | Your own score: a key (quality by default, or a new one), thumbs or 1 to 5, a comment and a correction. Scoring the same key again replaces it |
| Run evaluator | Scores the reply now with one of your model judges or code checks, and saves it as that evaluator's score. Judges run on your LLM key; 30 a minute |
| Add to dataset | Opens the Add to dataset sheet with the question and this reply as the reference answer. Edit the answer first if the reply was wrong |
| Send to review | Puts the reply in one of your review queues |
| Open trace | Opens the reply's trace in Tracing, with every step it took |
None of them costs credits. See Feedback and scores, Datasets and examples and Review queues.
You're done when
You can find a conversation with the filters, open it, and a score you add shows on the reply and in its Feedback panel.
If something goes wrong
My Playground chats aren't listed
The list shows Live traffic by default. Choose Test or All, or Show test chats when the list is empty.
Open trace is greyed out
"This reply has no trace": SentientOne couldn't find one for it, which can happen for older replies.
Run evaluator says the agent has no LLM key
A judge set to Same as the agent borrows the agent's model and key. Add a key to the agent, or choose a judge model on the evaluator. A workflow's conversation always needs an evaluator with its own judge model.
Still stuck?
Email us with what you tried and any request id. A person replies.