CloudWatch Omni connects an agent's answer with the work behind it

The agent says it could not look up an order. Its request is right there on the screen. The service that could explain what happened is in another account, and its records are nowhere in the current view. This hypothetical investigation is a useful way to approach CloudWatch Omni: it can connect more of the work in one interface, while the location of each record still matters.

Omni brings together service discovery, traces of requests, evaluations of agent behavior and queries written in everyday language. AWS announced general availability on September 23, 2026, initially in US East (N. Virginia), US West (Oregon) and Europe (Ireland). Teams can use a standalone web experience with single sign-on or an extension for VS Code, Cursor and Kiro. AWS says local agent development through that extension needs no AWS account.

A shared sign-in does not move the records

After signing in, the investigator enters a space associated with an account and Region. The documentation calls the shared entry point and identity connection a domain. It gives the team one entrance, not automatic access to a single collection of everything behind it. CloudWatch centralization rules govern which telemetry, the measurements and records emitted by systems, moves between accounts and Regions.

Existing CloudWatch records remain useful, and the familiar console, alarms and APIs stay in place. Omni places agent evaluations beside traces, latency and token usage. The trace follows a request's steps and timing; the evaluation asks whether the answer or behavior met a chosen criterion. A poor answer beside its trace gives the investigator a way to follow what produced it, as long as the relevant service records are available there too.

The order service has the other half of the story

In our imagined case, the agent's steps are visible but the order service's logs never reached this space. An empty search does not establish whether that service received the request. It tells the investigator that part of the route is still missing. Finding it through authorized access is more useful than changing the agent or repeating consequential work because no record appeared.

A small trial can expose that boundary. Send one synthetic operation with a known downstream result, then follow its identifiers through traces and logs to the same service and time window. Deliberately omit one source and see whether the responder notices the gap. Our OpenTelemetry explanation develops this connection: records become useful when someone can tie them to the operation being investigated.

If policy prevents bringing the order-service records into the space, a person can bridge the investigation instead. Identify someone authorized to inspect the other account and give them the request ID. They can find the missing part without moving every record or expanding anybody's permissions. That handoff is part of the working route, not a failure to put everything on one screen.

Price the investigation the team will actually use

AWS separates ingestion, storage and analysis charges, and agent evaluations use AgentCore Evaluations rates. Compare those recurring costs with the team's current investigation process. A documented collection of features does not establish a measured reduction in resolution time.

For this order lookup, the worthwhile improvement is specific. Start with the failed agent call, find the matching service record or the person allowed to retrieve it, and explain what happened. A trial that completes that route gives the team something concrete to weigh against the cost of keeping it available.