Key Takeaways

  • Amazon CloudWatch Omni connects AI agent traces with application and infrastructure telemetry in a standalone experience.
  • Evaluation workflows help teams diagnose agent behavior, compare changes, and monitor production quality.
  • OpenTelemetry support and Azure ingestion broaden Omni beyond AWS-native workloads, though deeper multi-cloud coverage is still ahead.

Amazon Web Services has launched Amazon CloudWatch Omni, an AI-powered observability experience designed to help development and operations teams investigate AI agents alongside the applications and infrastructure supporting them.

The launch reflects a practical shift in enterprise AI. Agents are moving beyond demonstrations into support, infrastructure management, expense approvals, software delivery, and other business workflows. IDC expects more than 1 billion AI agents to be actively deployed by 2029, collectively carrying out roughly 217 billion actions each day. That scale raises a difficult operational question: How does a business know that an agent did the right thing, rather than merely completing a task without triggering a conventional error?

Traditional monitoring does not fully answer that question. An application might report healthy latency, availability, and error rates while an agent selects the wrong tool, retrieves stale information, produces a toxic response, or follows an unhelpful reasoning path. Agent behavior can also vary between invocations, even when the input appears similar.

CloudWatch Omni approaches this problem by placing prompts, model calls, tool invocations, sub-agent interactions, logs, metrics, and traces within a shared investigation environment. Operators can follow an incident from an agent decision into an application dependency and then down to the underlying infrastructure. A questionable answer, for example, might originate with poor reasoning. Or it could result from an API failure caused by an exhausted database connection pool. Those are very different fixes.

Observability teams have spent years assembling separate products for application performance monitoring, infrastructure health, logging, tracing, and incident response. Generative AI has added another layer, complete with evaluations, prompt experiments, model configuration, and retrieval analysis. Amazon is positioning Omni as the connective tissue rather than another isolated console.

The product runs outside the AWS Console, using a dedicated web interface for operators and extensions for VS Code, Cursor, and Kiro. Developers can inspect traces while working locally, while production teams can use enterprise SSO and role-based permissions in the web experience. Both interfaces operate on the same CloudWatch data layer.

The agent evaluation workflow connects diagnostic and monitoring capabilities. Teams diagnose a flagged trace, score the output for characteristics such as correctness and retrieval quality, compare a revised configuration against production-derived datasets, and then monitor the selected version against live traffic. Evaluators can examine coherence, tool selection, guardrail enforcement, and memory behavior.

Amazon CloudWatch Omni supports CrewAI, LangGraph, OpenAI Agents SDK, Strands, and Vercel AI SDK. Amazon Bedrock AgentCore users receive native telemetry when observability is enabled, along with built-in evaluators. Existing tools are not necessarily displaced: AWS says evaluations from Braintrust, DeepEval, and Ragas can also be brought into Omni.

Open standards are an important part of that pitch. The OpenTelemetry project provides a vendor-neutral framework for collecting and exporting traces, metrics, and logs, and Omni accepts telemetry through OTLP endpoints. That gives teams a route into the service without requiring every workload to use AWS-specific instrumentation.

For existing CloudWatch customers, AWS says current logs, metrics, traces, dashboards, and alarms carry into Omni without reconfiguration. New users create an Omni space, send OTLP telemetry, and allow the service to discover topology and dependencies. Natural-language and SQL queries span different signal types, while AWS DevOps Agent assists with anomaly detection, hypothesis generation, and root-cause investigation. Slack and GitHub integrations can route alerts into existing team workflows.

CloudWatch Omni already ingests application and agent telemetry from Azure, with broader multi-cloud capabilities planned. That matters, but buyers will likely examine how much operational depth Omni offers outside AWS, not simply whether it can ingest the data.

Pricing is tied primarily to telemetry sent and stored. Dashboards and alerts carry no additional charge, query volume up to 5 times monthly ingestion is included, and eligible accounts receive a 30-day trial plus $1,000 in credits for OpenTelemetry ingestion. Capital One participated as a design partner, while Sony Group Corporation has tested Omni across an enterprise agentic AI environment supporting hundreds of proof-of-concept and production workloads.

The larger bet is that agent evaluation and conventional observability are becoming one discipline. If Amazon can reduce the distance between a bad AI outcome, its technical cause, and a validated fix, CloudWatch Omni could give AWS a stronger position as enterprises turn experimental agents into operational systems.