AWS’s new observability experience connects agent evaluations with application telemetry, but teams still need to manage data flows, access and costs.
AWS is adding a shared observability workspace for applications and AI agents, putting agent traces and quality evaluations alongside operational signals such as errors and latency. The company published launch posts on September 22, 2026; its release note lists CloudWatch Omni’s general availability as September 23. The distinction matters: September 22 was the announcement date, while September 23 is the date AWS gives for GA. (AWS News Blog, Sept. 22; AWS release note, Sept. 23)
The launch is aimed at a practical problem for teams putting AI agents into production: a service can appear healthy by conventional measures while its agent is returning poor answers or mishandling tool calls. Omni gives teams a place to examine those quality signals alongside application telemetry. It also adds natural-language investigations and a standalone web experience, rather than requiring engineers to work only inside the AWS Management Console.
What changed in CloudWatch Omni
Omni is a new experience built on CloudWatch, not a replacement for the existing service. AWS says current CloudWatch dashboards, alarms, APIs and console workflows continue to work. Existing telemetry already sent to CloudWatch can appear in Omni without re-instrumentation; workloads instrumented with OpenTelemetry can send data using OTLP. (AWS CloudWatch Omni documentation)
The interface groups telemetry around applications and services, including automatically discovered dependencies. Engineers can ask questions in plain language; the built-in assistant can generate SQL or PromQL queries against available telemetry. For agent workloads, teams can evaluate traces and inspect quality signals alongside operational measures. AWS says the agent observability workflow supports multiple frameworks, including OpenAI Agents SDK, LangGraph, CrewAI, Vercel AI SDK and Strands. (AWS release note; AWS documentation)
The service is generally available in US East (N. Virginia), US West (Oregon) and Europe (Ireland), according to AWS. Its web experience uses organization-level sign-in, and AWS also lists IDE extensions for VS Code, Kiro and Cursor. Availability is therefore a deployment consideration: teams operating elsewhere should confirm regional support before designing around Omni. (AWS release note)
Why agent teams may care
Traditional observability helps answer whether a service is slow, unavailable or generating errors. Agent systems introduce a different failure mode: a request can complete normally while the model misunderstands the task, chooses the wrong tool or produces an incorrect response. Evaluating agent traces alongside latency and errors could help teams connect a user-facing quality regression to the execution path that produced it.
That is the product’s clearest operational promise—not that AI can replace an on-call engineer, but that teams may have fewer separate places to consult when diagnosing a failure. This is an inference from the combined workflow AWS describes, not a demonstrated performance result: the announcement does not establish that Omni will resolve incidents faster than a team’s existing tools.
OpenTelemetry support may also ease adoption for teams that already instrument services with standard telemetry. OTLP is a defined protocol for sending logs, metrics and traces between sources and backends; using it can reduce dependence on a vendor-specific instrumentation path. But a common protocol does not make every backend feature interchangeable. Teams should treat their instrumentation and their observability platform as separate choices. (OpenTelemetry Protocol specification)
The caveats: centralization, spend and sensitive data
Omni does not centralize telemetry across AWS accounts or Regions by itself. AWS’s documentation says teams must configure CloudWatch centralization for that data to appear together. Organizations with fragmented accounts should include that setup—and its associated data movement—in their evaluation, rather than assuming the new interface consolidates everything automatically. (CloudWatch Omni documentation)
The service also adds usage dimensions to assess. AWS’s pricing page describes charges for telemetry ingestion and storage, as well as analysis; agent evaluations use Amazon Bedrock AgentCore Evaluations rates. Costs will depend on a team’s data volume, retention, query patterns and evaluation frequency, so “turn it on and see” is not a reliable budget strategy for a high-volume production system. (CloudWatch Omni pricing)
Agent traces can contain prompts, responses, retrieved documents and personal information. AWS says Omni does not automatically detect or redact personally identifiable information; teams must configure filtering or redaction at an appropriate point in their pipeline. AWS also documents cross-Region processing for some AI features. It says Omni content is not used to train foundation models, but teams should still review what telemetry they send, which models they configure and where requests may be processed. (AWS guidance on sensitive data; cross-Region inference documentation; data-use policy)
What teams should do now
CloudWatch users running agents in a supported Region can test Omni against one non-critical service or workflow. Before routing production traces, inventory which signals the team needs, configure sensitive-data filtering, confirm account and Region centralization, and model expected ingestion, storage and evaluation costs. Then compare the assistant’s generated queries and evaluations with existing incident procedures; keep a human review step for consequential actions.
For teams not already invested in CloudWatch, Omni is a reason to evaluate the workflow—not, by itself, a reason to move observability platforms. AWS has announced the feature set and pricing dimensions, but teams still need to establish how useful its investigations are on their own workloads, how its costs compare with their current stack, and whether its regional availability fits their deployment. The practical takeaway is to treat Omni as a new way to inspect telemetry, not as a substitute for good instrumentation, data controls or incident judgment.