Skip to main content

Traces

Availability

Open core · self-host + all Splyntra Cloud plans

A trace is one agent run — the hierarchical timeline of everything the agent did: planning steps, LLM calls, tool invocations, retrievals, and sub-agent handoffs, each carrying latency, token counts, and cost. Traces are the first-class unit in Splyntra.

The explorer​

The Traces screen lists runs, newest first, with duration, cost, token totals, and risk score per run. Like every table in the dashboard it is searchable, sortable, and paginated, with Excel export.

Filter to narrow in on what matters:

  • Project and environment
  • Agent / service
  • Status (success vs. error)
  • Risk (for example, only runs above a risk threshold)
  • Time range

The trace-detail waterfall​

Open a run to see the waterfall — the full span tree laid out on a timeline. Each span shows its type (agent, step, llm_call, tool_call, retrieval, vector_search), duration, token counts, and cost, so you can spot the slow model call, the retry, or the tool that failed. Selecting a span reveals its attributes and its (redacted) input/output.

Risk drill-down​

Every run carries a single 0–100 risk score built from secret/PII detection, prompt-injection detection, and content moderation, mapped down onto the exact spans where the risk appeared. From a run's score you can jump straight to the offending span — no separate SIEM, no correlation guesswork. The span highlights the finding (for example, a leaked API key in a tool call's arguments).

See The risk score for how the score is built, and Security for the incident views.

Next steps​