AWS Combines AgentCore and DevOps Agent to Unify Multi-Agent Telemetry and Online Evaluation

On September 11, 2026, AWS published an operational architecture uniting Amazon Bedrock AgentCore Observability, AgentCore Evaluations, and the AWS DevOps Agent to manage the production lifecycle of multi-agent AI swarms. The implementation utilizes OpenTelemetry instrumentation inside the AgentCore runtime to emit distributed traces, execution metrics, and asynchronous online evaluation scores directly into Amazon CloudWatch, allowing AWS DevOps Agent to autonomously investigate incidents using unified operational and evaluation telemetry. Traditional infrastructure monitoring fails to capture the failure modes unique to agentic architectures. In a multi-agent system, an identity and access management misconfiguration can cause an agent invocation to return an empty payload without generating an HTTP 500 status code, or an ambiguous prompt in a supervisor agent can route 20 percent of user requests to the wrong specialist while resource utilization and error rates stay green. By pairing OpenTelemetry-based runtime tracing with asynchronous online evaluations—sampling between 0.01 and 100 percent of live traffic—teams can detect behavioral degradation, hallucination rates, and task routing failures without introducing latency to user-facing paths. This release reflects the broader convergence of distributed tracing and AI operational telemetry across modern cloud platforms. As organizations transition from single-prompt LLM wrappers to autonomous multi-agent swarms using patterns like Swarm, Graph, and Agents-as-Tools, observability vendors are racing to treat LLM spans, tool calls, and model evaluation metrics as first-class citizens. By standardizing the ingestion layer on OpenTelemetry and funneling telemetry into automated triage workflows via DevOps agents, AWS is reinforcing the industry shift where observability platforms serve as both diagnostic datastores and direct input layers for autonomous remediation. In practice, engineering teams deploying agent workflows should decouple infrastructure uptime from agent task efficacy. Observability architectures must incorporate dual-layer monitoring: baseline telemetry for container and API health, paired with continuous asynchronous scoring for semantic accuracy and tool-selection fidelity. SRE teams should configure configurable trace sampling to balance cost and coverage in high-throughput environments, while ensuring that automated triage agents have programmatic access to correlated CloudWatch spans and evaluation outputs to quickly pinpoint whether failures originate from prompt degradation, API dependencies, or orchestration logic.