CloudWatch Omni brings agent evaluation into the telemetry view

A failed agent task can leave evidence in several places: the model response, the tool call and the service that received it. AWS’s new CloudWatch Omni brings those investigations into a shared interface. The useful adoption test is whether a responder can follow the same operation across those records, including the account and Region where the work happened.

AWS announced general availability on September 23, 2026, initially in US East (N. Virginia), US West (Oregon) and Europe (Ireland). Omni provides a standalone web experience with SSO and an IDE extension for VS Code, Cursor and Kiro. AWS says the extension supports local agent development without an AWS account.

What changes for an investigation

The Omni documentation describes service discovery, natural-language query generation and agent evaluations alongside traces, latency and token usage. Existing CloudWatch telemetry remains usable; Omni does not replace the existing console, alarms or APIs. These are documented capabilities, not independently measured improvements in incident resolution.

There is an important scope distinction. A domain provides the entry point and identity connection. A space is associated with an account and Region. Omni does not independently centralize telemetry across accounts or Regions; that still depends on CloudWatch centralization rules. A shared sign-in therefore does not, by itself, establish that the records needed for an investigation are present.

Test one operation across the boundary

Consider a hypothetical agent that reports an order lookup failed. Its own spans are visible, but the order service runs in another account whose telemetry has not reached the selected space. An empty query result cannot establish that the service never received the request. Before changing the agent or retrying consequential work, inspect the destination’s records through an authorized path.

For a pilot, choose a synthetic operation with a known downstream result. Check that its trace and log identifiers lead to the same service and time window, then repeat with a deliberately missing telemetry source. Record whether the responder can identify the gap and choose the next investigation step. This applies the evidence-continuity principle in our OpenTelemetry explanation to Omni’s account and Region boundaries.

If policy prevents centralizing a source, keep the investigation distributed and document the authorized handoff. Do not broaden data access merely to complete a dashboard. The acceptance criterion becomes a working evidence handoff, rather than every record appearing in one space.

Include the cost of the evidence

AWS lists separate ingestion, storage and analysis charges, with agent evaluations billed at AgentCore Evaluations rates. Include those costs when comparing the proposed investigation path with today’s process. The pilot should establish what decision becomes easier, which evidence remains unavailable and what it costs to maintain that improvement.