AWS AgentCore Observability pinpoints slow agents with CloudWatch tracing

Amazon Web Services has expanded Amazon Bedrock AgentCore with an Observability feature designed for the production phase of AI agents—the point where they work correctly but start degrading in performance. This second installment in AWS's series shifts focus from broken agents (infinite loops, tool errors) to slow ones. Using AgentCore Observability alongside Amazon CloudWatch Transaction Search, developers can now trace every request from invocation to response, identify where latency accumulates, and catch issues before users notice.

The methodology is concrete. First, query CloudWatch for agent invocations that exceed your performance budget—for example, a filter for Latency > 3000ms. Pick a representative RequestId, then examine its full OpenTelemetry trace. AWS shows a real case where three sequential execute_event_loop_cycle operations produced 17 spans with 7.5-8.2 seconds average latency, caused by tools like customer_lookup and order_history running in series rather than in parallel. The system also tracks memory retrieval latency per namespace, with a target of under 200ms—beyond that, users perceive delays. For long-running sessions, the observability tooling monitors memory growth to prevent unbounded expansion in production. This gives teams a repeatable workflow: set budgets, query for outliers, drill into spans, and optimize either tool parallelism or memory organization.