Your AI agent is a black box. Three layers of observability can

If your AI agent is in production, you cannot explain what it did, which tools it called, or how many reasoning cycles it took. Traditional observability monitors CPU and RAM, not agent reasoning. This gap leaves engineering teams blind to agent behavior and inefficiency. The video demonstrates three layers: metrics, traces, and attributes, all built

using open telemetry. Metrics reveal the agent's shape: total cycles, duration, token usage, and per-tool timings. This transforms a slow agent from a mystery into a diagnosis. Traces show the step-by-step sequence of decisions with full context, including model calls, tool executions, and intermediate outputs. This visibility identifies problems like repeated tool calls that produce

correct but inefficient answers. Attributes allow teams to add business context like booking thresholds directly into traces. This data lives in the telemetry, not the conversation, making it searchable months later. The demo uses a strand SDK that emits open telemetry data natively, requiring zero additional setup. In production, the same data flows to Amazon

Bedrock Agent Core via one configuration change. The approach claims to solve agent observability with minimal configuration, but the strength of evidence is mixed. The demo is clear and the architecture is sound. However, the video is essentially a product walkthrough for a strand and AWS services. The real insight is that open telemetry provides