SentientOneSentientOne Docsv1.0.1686
Back to Platform

Tracing

Step-by-step execution timeline for every API call — LLM call, MCP tools, knowledge retrieval, web fetch. The deepest debugging surface in the platform.

What you see

Tracing · app.sentientone.ai/trace
Trace · req_8b21d4f0Agent: AllRange: 24hTrace ID
14:32:08Website Chatbot · gpt-4o200 OK1.4 s
14:31:52Order Status Agent · claude-3.5200 OK2.1 s
14:31:14HR Assistant · claude-3.5error0.9 s
14:30:48Customer Support · gpt-4o-mini200 OK1.0 s

Latency

1.84 s

Prompt

824 tok

Completion

152 tok

Cost

$0.0042

LLM (62%)MCP get_order (22%)Retrieval (11%)Serialize (5%)
Execution flow

Request list

  • RowOne API call per row. Newest first. Columns: time, agent, model, status, duration, Trace ID.
  • Status pillNeutral for success, accent for error. Click any error row to jump straight into the failing step.
  • FiltersAgent, status, date range, and free-text search across the Trace ID. Filter chips combine.
  • SortDefault sort is newest first. Click Duration to surface your slowest requests.

Inspector

  • HeaderAgent name, model, status, and Trace ID. The Trace ID matches the X-Trace-Id header your server saw — copy it to correlate with your application logs.
  • Stat cardsTotal latency, prompt tokens, completion tokens, and estimated cost — the same fields you'd find on the Analytics breakdown table, framed for a single request.
  • Latency breakdown barColoured segments — LLM call, MCP tool, knowledge retrieval, web fetch. Hover any segment for its duration in milliseconds. Mostly-LLM bars indicate model-bound latency; mostly-tool bars indicate slow integrations.
  • MCP ToolsEvery MCP server + tool the agent had access to during this call. If you expected a specific tool to fire and it's not listed here, the agent chose not to call it (check the system prompt) or the server wasn't enabled at request time.
  • Execution FlowNested tree of every recorded step with model, latency, status, and payload metadata. Click a node to expand its fields. Failed steps are highlighted so you can find the breakage without scrolling.

Debug a failing request

  1. 1

    Find the request

    Filter the request list by the failing agent and a status of error. Sort by time to find the most recent failure.
  2. 2

    Open the inspector

    Click the row. Scan the latency breakdown bar — the colour that dominates points at which subsystem caused the issue.
  3. 3

    Drill into Execution Flow

    Expand the failing step (it's highlighted). The error_message field carries the root cause and the upstream provider's response code.
  4. 4

    Copy the Trace ID

    Paste it into your application logs (or your Sentry dashboard) to correlate with the upstream caller that triggered the request.

Retention

Trace retention follows your plan — 14 days on Starter, 90 days on Pro, 365 days on Enterprise. Older traces are aggregated into Analytics indefinitely.