Skip to main content

Telemetry Configuration

Enable distributed tracing to monitor agent behavior, debug issues, and track performance using OpenTelemetry.

Complete Reference

For complete field documentation, backend setup, and collector configuration, see agent.yml → Telemetry.

Overview​

Telemetry provides visibility into your agent's operations through distributed tracing. When enabled, Dexto automatically traces agent operations, LLM calls, and tool executions.

What you get:

  • Complete request lifecycle traces
  • LLM token usage tracking
  • Tool execution monitoring
  • Export to any OTLP-compatible backend

Quick Start​

1. Start Jaeger (Local)​

docker run -d \
  --name jaeger \
  -p 16686:16686 \
  -p 4318:4318 \
  jaegertracing/all-in-one:latest

2. Configure Agent​

telemetry:
  enabled: true
  serviceName: my-agent
  export:
    type: otlp
    endpoint: http://localhost:4318/v1/traces

3. View Traces​

Open http://localhost:16686 and explore your traces.

Configuration Options​

telemetry:
  enabled: boolean              # Turn on/off (default: false)
  serviceName: string           # Service identifier in traces
  tracerName: string            # Tracer name (default: 'dexto-tracer')
  export:
    type: 'otlp' | 'console'    # Export destination
    protocol: 'http' | 'grpc'   # OTLP protocol (default: 'http')
    endpoint: string            # Backend URL
    headers:                    # Optional auth headers
      [key: string]: string

Export Types​

OTLP (Production)​

Export to OTLP-compatible backends:

telemetry:
  enabled: true
  serviceName: my-prod-agent
  export:
    type: otlp
    endpoint: http://localhost:4318/v1/traces

Console (Development)​

Print traces to terminal:

telemetry:
  enabled: true
  export:
    type: console

Common Configurations​

Local Jaeger​

telemetry:
  enabled: true
  serviceName: my-dev-agent
  export:
    type: otlp
    protocol: http
    endpoint: http://localhost:4318/v1/traces

Grafana Cloud​

telemetry:
  enabled: true
  serviceName: my-prod-agent
  export:
    type: otlp
    endpoint: https://otlp-gateway-prod.grafana.net/otlp
    headers:
      authorization: "Basic $GRAFANA_CLOUD_TOKEN"

Honeycomb​

telemetry:
  enabled: true
  serviceName: my-prod-agent
  export:
    type: otlp
    endpoint: https://api.honeycomb.io:443
    headers:
      x-honeycomb-team: $HONEYCOMB_API_KEY

What Gets Traced​

Dexto automatically traces:

  • Agent operations - Full request lifecycle
  • LLM calls - Model invocations with token counts
  • Tool executions - Tool calls and results

Key attributes:

  • gen_ai.usage.input_tokens - Prompt tokens
  • gen_ai.usage.output_tokens - Completion tokens
  • llm.provider - Provider name
  • llm.model - Model identifier

Use Cases​

ScenarioHow Telemetry Helps
Debug slow requestsIdentify bottlenecks in traces
Monitor token usageTrack LLM costs and optimize prompts
Production monitoringSet alerts for errors and latency
Performance optimizationFind inefficient operations

Performance Impact​

Minimal overhead:

  • ~1-2ms per span
  • Async export (non-blocking)
  • Automatic batching

For high-volume agents, consider sampling or using a collector.

Best Practices​

  1. Enable in production - Essential for observability
  2. Use meaningful service names - Different names per deployment
  3. Set up monitoring - Create alerts for issues
  4. Consider sampling - For high-traffic scenarios
  5. Use collectors - For advanced processing and buffering

See Also​