AWS puts the trace waterfall inside the tool response, betting verification is the slow part

Two halves arrive in one response, and they are not the same kind of artifact. The text summary is model output. The picture, according to AWS, is produced by executing code against the same data sources that power your dashboards, which is why the company describes the result as deterministic and says you are not trusting the AI's interpretation ^[6]. That distinction is narrower than it sounds. The render settles whether a number really came out of the store. It does not settle whether the agent chose the right trace, the right window, or the right service, and the post does not claim it does ^[12].

The arithmetic is worth doing on AWS's own description of the current loop. It has five human steps: ask the agent, leave the IDE and log in to a separate observability UI, re-run the queries manually, compare the text against the dashboards, then go back and pick up where you left off ^[4]. Three of those are mechanical, and those three are what an in-thread widget can absorb ^[11]. Asking and resuming stay, along with the judgement call about whether the hypothesis holds. AWS is candid that the agent already saved the query time and saved none of the tab-switching ^[2], and it names that external verification loop as the bottleneck that pushes engineers into a tool-switching role ^[3].

Where this gets interesting operationally is the delivery path. The flow starts in an IDE or AI desktop client, with Claude, VS Code and Cursor listed ^[9], and the visualization has to be drawn in that client's chat window ^[1]. The useful part of the feature therefore depends on software AWS does not ship. A host that shows only the text half leaves you with the same loop described above.

Underneath, a local MCP server runs on the engineer's machine and brokers authenticated queries to the OpenSearch UI endpoint ^[7]. That endpoint is the serverless front for OpenSearch domains, serverless collections, CloudWatch and Amazon Managed Service for Prometheus ^[8]. One process on a laptop, with a read path across four telemetry backends, is a credential question before it is a productivity question, and the excerpt does not go further than calling the server a secure bridge ^[7].

AWS is unusually direct about who this is for: teams that ran agentic observability locally for control and cost efficiency, took a hit on ease of use and sometimes on agent performance, and now carry verification as their main operational burden ^[10]. The offer to them is ergonomics rather than a cleverer agent. That is a smaller claim than this category usually makes, and a testable one. Either the browser tab stops opening during incidents, or it does not.