Collapse navigation tree

Collapse

View Transcripts

Redpanda Agentic Data Plane is supported on BYOC clusters running with AWS and Redpanda version 25.3 and later. It is currently in a limited availability release.

Use the Transcripts view to filter, inspect, and debug agent execution records. Filter by operation type, time range, or service to isolate specific executions, then drill into span hierarchies to trace request flow and identify where failures or performance bottlenecks occur.

For conceptual background on spans and trace structure, see Transcripts and AI Observability.

After reading this page, you will be able to:

  • Filter transcripts to find specific execution traces

  • Use the timeline interactively to navigate to specific time periods

  • Navigate between detail views to inspect span information at different levels

Prerequisites

  • Running agent or MCP server with at least one execution

  • Access to the Transcripts view (requires appropriate permissions to read the redpanda.otel_traces topic)

Filter transcripts

Use filters to narrow down transcripts and quickly locate specific executions. When you use any of the filters, the transcript table updates to show only matching results.

The Transcripts view provides several quick-filter buttons:

  • Service: Isolate operations from a particular component in your agentic data plane (agents, MCP servers, or AI Gateway)

  • LLM Calls: Inspect large language model (LLM) invocations, including chat completions and embeddings

  • Tool Calls: View tool executions by agents

  • Agent Spans: Inspect agent invocation and reasoning

  • Errors Only: Filter for failed operations or errors

  • Slow (>5s): Isolate operations that exceeded five seconds in duration, useful for performance investigation

You can combine multiple filters to narrow results further. For example, use Tool Calls and Errors Only together to investigate failed tool executions.

Toggle Full traces on to see the complete execution context, in grayed-out text, for the filtered transcripts in the table.

Filter by attribute

Click the Attribute button to query exact matches on specific span metadata such as the following:

  • Agent names

  • LLM model names, for example, gemini-3-flash-preview

  • Tool names

  • Span and trace IDs

You can add multiple attribute filters to refine results.

Use the interactive timeline

Use the timeline visualization to quickly identify when errors began or patterns changed, and navigate directly to transcripts from specific time windows when investigating issues that occurred at known times.

Click on any bar in the timeline to zoom into transcripts from that specific time period. The transcript table automatically scrolls to show operations from the time bucket in view.

When viewing time ranges with many transcripts (hundreds or thousands), the table displays a subset of the data to maintain performance and usability. The timeline bar indicates the actual time range of data currently loaded into view, which may be narrower than your selected time range.

Refer to the timeline header to check the exact range and count of visible transcripts, for example, "Showing 100 of 299 transcripts from 13:17 to 15:16".

Inspect span details

The transcript table shows:

  • Time: When the span started (sortable)

  • Span: Span type and name with hierarchical tree structure

  • Duration: Total time or relative duration shown as visual bars

To view nested operations, expand any parent span. To learn more about span hierarchies and cross-service traces, see Transcripts and AI Observability.

Click any span to view details in the panel:

  • Summary tab: High-level overview with token usage, operation counts, and conversation history.

  • Attributes tab: Structured metadata for debugging (see standard attributes by layer).

  • Raw data tab: Complete OpenTelemetry span in JSON format. You can also view raw transcript data in the redpanda.otel_traces topic.

Rows labeled "awaiting root — waiting for parent span" indicate incomplete traces. This occurs when child spans arrive before parent spans due to network latency or service failures. Consistent "awaiting root" entries suggest instrumentation issues.

Common investigation tasks

The following patterns demonstrate how to use the Transcripts view for understanding and troubleshooting your agentic systems.

Debug errors

  1. Use Errors Only to filter for failed operations, or review the timeline to identify and zoom in to when errors began occurring.

  2. Expand error spans to examine the failure context.

  3. Check preceding tool call arguments and LLM responses for root cause.

Investigate performance issues

  1. Use the Slow (>5s) filter to identify operations with high latency.

  2. Expand slow spans to identify bottlenecks in the execution tree.

  3. Compare duration bars across similar operations to spot anomalies.

Analyze tool usage

  1. Apply the Tool Calls filter and optionally use the Attribute filter to focus on a specific tool.

  2. Review tool execution frequency in the timeline.

  3. Click individual tool call spans to inspect arguments and responses.

    1. Check the Description field to understand tool invocation context.

    2. Use the Arguments field to verify correct parameter passing.

Monitor LLM interactions

  1. Click LLM Calls to focus on model invocations and optionally filter by model name and provider using the Attribute filter.

  2. Review token usage patterns across different time periods.

  3. Examine conversation history to understand model behavior.

  4. Spot unexpected model calls or token consumption spikes.

Trace multi-service operations

  1. Locate the parent agent or gateway span in the transcript table.

  2. Use the Attribute filter to follow the trace ID through agent and MCP server boundaries.

  3. Expand the transcript tree to reveal child spans across services.

  4. Review durations to understand where latency occurs in distributed calls.

Next steps