Observability patterns for AI agents on Amazon EKS

This article demonstrates two observability patterns for AI agents running on Amazon EKS, addressing the challenge of monitoring non-deterministic agent behavior, token consumption, and costs at scale.

• Decentralized pattern: Agents export traces directly to Langfuse with no shared collector, providing isolation and simplicity for small deployments
• Centralized pattern: Traces flow through a shared OpenTelemetry Collector for batching, sampling, and multi-backend routing without agent code changes
• Both patterns use Strands SDK auto-instrumentation, Bifrost LLM gateway for Amazon Bedrock routing, and OpenTelemetry GenAI semantic conventions
• Trace correlation via W3C traceparent headers enables unified visibility across agent invocations, tool calls, and model inference with per-span token counts and costs
• Decentralized suits fast setup and small teams; centralized scales better with fleet-wide control and operational consistency
• Complete implementation provided with deployment scripts, agents, gateway, and observability stack running on Amazon EKS Auto Mode

Both patterns produce identical traces in Langfuse, allowing teams to start decentralized and migrate to centralized as their agent fleet grows without code changes.