Conversations

Conversations is where you find out what your agent is actually like to use. Every exchange is stored with its transcript and the runs that produced it, so you can read a real answer instead of guessing how the prompt behaved.

It is also the highest-value page for improving the agent. Aggregate metrics tell you something changed; transcripts tell you what to fix.

Statuses

A conversation's status reflects its most recent run.

Filtering to Failed is the quickest way to find the conversations worth reading first.

Filters

• Search — match on conversation title or ID
• Date range — presets or a custom range up to three months back
• Status — any combination of the five statuses
• User — narrow to one person, shown as a chip you can clear

Filters live in the URL, so a filtered view is a link you can paste into a ticket or a channel.

Export conversations

Choose Export to download every conversation matching the current search, date range, status, and user filters. The export includes all matching pages, always includes dashboard preview traffic, and offers three detail levels:

JSON exports use nested messages and runs arrays. Summary CSV is one file. Transcript CSV downloads a ZIP containing conversations.csv and messages.csv; Full diagnostics adds runs.csv. The files join on conversation IDs.

Internal tenant and end-user keys, raw usage payloads, and other restricted fields are never included. A synchronous export can contain up to 10,000 conversations or 25 MiB; use a narrower date range, filters, or detail level if either limit is reached.

Conversation detail

The drawer opens with the conversation's ID, user, status, and start and last-activity timestamps, then two sections.

Transcript opens in rendered view, reproducing the user-visible conversation: formatted Markdown, files, generated components, sources, and collapsed activity. It intentionally leaves out the composer, response actions, avatars, timestamps, and other Agent controls so you can focus on the output itself. Safe interactions inside the output still work, but the preview cannot retry a response, submit feedback or approvals, invoke client tools, or trigger generated-component actions.

Turn on View raw to inspect the literal emitted text with Markdown syntax, message bubbles, and timestamps. Structured parts are summarized instead of rendered in this view. Your choice lasts only while the monitoring page remains open and is not added to the URL or saved in the browser.

Both views update from the same live transcript while the conversation is streaming.

Runs lists the model calls behind the conversation, each with its model, duration, and status, linking into Logs. A single message can involve several runs when the agent calls tools before answering.

Using transcripts to improve the agent

Reading transcripts in bulk surfaces patterns no metric will:

• Questions the agent could not answer are your knowledge base backlog, in priority order
• Answers that were technically right but unhelpful are usually a system prompt problem — tone, length, or missing escalation paths
• The agent explaining what it cannot do points at a missing tool
• Repeated rephrasing by the same person means the first answer missed, even though nothing failed

Change one thing at a time and read the same class of conversation again. Prompt changes apply to the next message, so the loop is short.

Next steps

• Tune the system prompt based on what you read
• Add knowledge documents for the questions that went unanswered
• Inspect run details for latency, tool activity, and errors