Conversation review

Inspect captured conversations: transcripts, operational timeline, traces, annotations, and automatic outcome analysis.

The Agent Playground tells you how an agent behaves when you ask. Conversation review tells you how it behaved for real users — and gives your team the evidence to improve prompts, tests, and releases.

Open conversations

  1. Enter a team workspace.
  2. In the sidebar, open Quality and choose Conversations under Review.
  3. Filter or search for a thread, then open the run detail view.
Conversation review with transcript, timeline events, and trace contextConversation review with transcript, timeline events, and trace context

What you see on a run

Each conversation thread opens a detail page with tabs:

TabPurpose
ConversationTranscript plus operational timeline — what was said and what the platform recorded during the run.
LineageInvocation tree and trace context across nested agent or tool calls.
AnalysisAutomatic outcome analysis when enabled for your workspace.
AnnotateRubric-based human review — score quality and capture findings.

Use Conversation for day-to-day debugging; use Lineage when you need to follow tool calls or nested invocations; use Annotate and Analysis for quality programs and regression tracking.

Typical review workflow

  1. Find — Filter conversations by agent, channel, environment, or time range.
  2. Read — Walk the transcript alongside timeline events (tool calls, handoffs, errors).
  3. Trace — Open Lineage when the failure is in execution, not wording.
  4. Score — Annotate with your team rubric or review automatic analysis results.
  5. Act — Turn findings into prompt changes, new test cases, or a held deployment.

When to use review vs the playground

SituationPrefer Agent PlaygroundPrefer Conversation review
Iterating on a prompt change✓
Reproducing a customer-reported issue✓
Auditing tool calls in production✓
Building a quality scorecard✓
Demo to stakeholders in the room✓

← MCP servers and tools · Testing and release →