agno-observability

Agno v2 is auto-instrumented with OpenTelemetry, so you only configure the exporter. In AgentOS set tracing=True; outside AgentOS initialize OTel manually and point OTEL_EXPORTER_OTLP_ENDPOINT at your backend, such as a local collector on localhost:4317.

X-Ray tracing of Agno spans requires the ADOT Lambda layer, since X-Ray does not accept OTLP natively. Attach the ADOT layer, set AWS_LAMBDA_EXEC_WRAPPER=/opt/otel-instrument and OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317, and grant AWSXRayDaemonWriteAccess.

No, Tracing: Active alone only captures infrastructure segments for Lambda and API Gateway, such as cold starts and invocation duration. The internal Agno spans for model calls, tools, and tokens require the ADOT layer to forward OTLP data to X-Ray.

Build a standard logging.Logger with a StreamHandler that emits one-line JSON, then pass it to configure_agno_logging via custom_default_logger and custom_agent_logger. Log one event per request with fields like request_id, provider, latency_ms, and used_fallback.

Production uses exclusively AWS managed services: CloudWatch Logs for logs, X-Ray via ADOT for traces, and CloudWatch EMF for metrics with alarms. SigNoz, Grafana, and Langfuse are local development backends only and should not receive production traffic.

High-cardinality dimensions like session_id or user_id explode metric costs and storage. Those identifiers belong in logs and traces for correlation, while metrics should aggregate low-cardinality dimensions such as provider or result.