Skip to main content
Polos provides built-in tracing for every agent and workflow execution. Every LLM call, tool invocation, and state transition is automatically captured - no additional instrumentation required.

​
Viewing traces

Open the Polos dashboard at http://localhost:5173 to view execution traces. Each agent or workflow run is recorded with full detail.

​
Timeline view

The timeline view shows every step of your agent’s execution in chronological order, including LLM calls, tool invocations, and their inputs/outputs: Polos dashboard showing agent execution trace

​
Graph view

The graph view visualizes the execution flow, making it easy to see how steps relate to each other and understand the structure of complex workflows: Polos dashboard showing agent graph

​
Understanding spans

Polos organizes traces into spans that represent units of active execution. A key concept is that spans end when an agent or workflow enters a waiting state - this reflects how Polos manages resources efficiently.

​
Example: Weather agent execution

Consider a weather agent that:
  1. Starts execution
  2. Calls the LLM
  3. LLM decides to call the get_weather tool
  4. Agent enters waiting state while tool executes
  5. Resumes after tool completes
  6. Calls LLM again with the tool result
  7. Returns final response
This execution produces two spans: Span 1: Initial execution
  • Agent starts
  • Calls LLM
  • LLM requests tool call
  • Agent enters waiting state (span ends)
Span 2: Resumed execution
  • Agent resumes after tool completes
  • Calls LLM with tool result
  • Agent returns final response (span ends)
This span structure reflects how Polos handles durability. When an agent enters a waiting state, it releases compute resources. The agent’s state is persisted, and when the wait condition is satisfied, a new span begins to continue execution. If a crash occurs between spans, the agent resumes from exactly where it left off. Spans during agent execution

​
What gets traced

Polos automatically captures:
  • LLM calls: Model, prompt, response, token usage, latency
  • Tool invocations: Tool name, inputs, outputs, duration
  • State transitions: When agents pause, resume, or complete
  • Workflow steps: Each step in a workflow with its result
  • Errors: Full error details and stack traces when failures occur

​
OpenTelemetry integration

Polos traces are compatible with OpenTelemetry. You can export traces to your preferred observability backend (Jaeger, Datadog, Honeycomb, etc.) for integration with your existing monitoring stack. More documentation on OpenTelemetry export configuration coming soon.