Traces

In development — not production-ready

Traces are still under development. The features on this page may change, and they are not yet ready for production use.

Explore what your agents actually did.

Trace list

One row per trace, with per-column filters — service (a dropdown of the agents present), span type, query, response, status, min tokens/latency, and context id — plus a rolling time window (from the last 5 minutes up to 24 hours) or an explicit date range. Hover a service to see its model and input/output token split. Select traces to add them to a dataset or run Detect Issues. Filters are remembered as you navigate away and back.

Context id: tie an agent to the tools it used

The Context ID column is the agent conversation id (gen_ai.conversation.id). It's what makes the trace list answer real questions:

• Which tools did this agent call? Because a trace's execute_tool spans nest under the invoke_agent span, filtering to a service — or clicking a tool from an alert or the Summary — shows exactly which agent turns invoked which tools, with error status.
• Follow a whole conversation. All turns sharing a context id belong to one session, and agent-to-agent hops keep the same id, so an orchestrator → sub-agent → tool flow reads as one connected trace. See why the conversation id matters.

Live tab

A real-time feed of spans as they're ingested — with instant client-side filtering (text, span type, OK/ERROR) and a rolling activity timeline. Great while wiring up instrumentation: fire a span and watch it arrive. See the Quickstart.

Trace flow

Open a trace to see its swimlane graph — Agent, LLM Chat, and MCP Tool lanes — with status-colored nodes positioned by sequence and connected by parent/child edges. Click a node for its token donut and full attribute list.

Span detail

Every span exposes its known fields (agent, tool, model, tokens, duration, status) plus the complete raw attribute set you sent — so you can confirm exactly what your instrumentation emitted.