Skip to content

Execution Traces

Inspect every reasoning step, tool call, and decision an agent makes during task execution.

Overview

Execution traces provide full transparency into how agents process tasks. Every LLM call, reasoning step, tool invocation, and error is logged with token counts and costs, giving you a complete audit trail.

Trace viewer showing execution timeline

Trace Timeline

Each trace displays a sequential timeline of steps:

Step TypeDescription
ReasoningThe agent's thought process and decision-making
Tool CallExternal tool or function invocation
Tool ResultResponse from a tool call
ErrorAny errors encountered during execution

Viewing Traces

You can access execution traces from three places:

  1. Task detail page — click on a completed task to see its trace
  2. Agent detail "Traces" tab — view all traces for a specific agent
  3. Cost dashboard — traces linked from the top tasks by cost view

Step Details

Click any step in the timeline to expand it:

Reasoning steps show the agent's internal thought process — what it considered, what it decided, and why.

Expanded reasoning step

Tool call steps show the function called, parameters passed, and the result returned.

Expanded tool call step

Cost Tracking

Each trace displays:

  • Token count per step — input and output tokens
  • Cost per step — USD cost based on the model's pricing
  • Total trace cost — aggregate cost for the entire execution
  • Duration — time taken from start to finish

Use Cases

  1. Debug agent behavior — see exactly why an agent made a specific decision or produced unexpected output
  2. Cost optimization — identify expensive reasoning loops or redundant tool calls and optimize prompts to reduce them
  3. Quality review — verify the agent followed instructions correctly before approving its output

Tips

  • Check traces when an agent produces unexpected results — the reasoning steps usually reveal the issue
  • Use trace cost data to identify agents that need prompt optimization
  • Compare traces across similar tasks to spot inconsistencies in agent behavior

Built by HelixNode