Amazon CloudWatch Omni Brings Tracing and Evals for Agentic Workloads Into the IDE
On September 22, AWS introduced Amazon CloudWatch Omni, an observability product built for generative AI and agentic applications. It is available as a free VS Code and Kiro extension and as a standalone web experience, not in the AWS Management Console. The extension needs no AWS account to start. AWS credentials are only needed to call Bedrock models or to use the optional Cloud Login for persistent storage and team sharing.
Omni records full execution traces of every LLM call, tool invocation, and reasoning step on a hierarchical timeline. A Compare mode shows two traces side by side for debugging regressions. It ships 17 built-in evaluators, covering coherence, helpfulness, faithfulness, and routing correctness, and can plug in third-party evaluators such as DeepEval. Batch experiments compare agent variants on evaluation scores, latency, and token usage. Other features include a prompt playground, prompt versioning, a multi-turn Session Explorer, and an Agent Topology view.
Instrumentation uses open standards, OpenInference and ADOT, with support for LangChain, LangGraph, CrewAI, the OpenAI SDK, Strands, and the Vercel AI SDK in Python and TypeScript, plus native Bedrock AgentCore integration. Because it is OpenTelemetry-based and starts locally, it competes directly with LangSmith- and Langfuse-style tooling for teams already on AWS.
Read more — AWS News Blog
Amazon EventBridge Enhanced Custom Event Buses Add Ordering, Dedup, and Org-Wide Sharing
On September 24, AWS introduced enhanced custom event buses for EventBridge, aimed at enterprise-scale event-driven systems. The most significant change is organisation-wide sharing: one bus can be used by every account in an AWS Organization, with no cross-account resource policies or bus-to-bus forwarding rules.
The new capabilities include:
- Ordered delivery: publishers set an
EventGroupId, and subscribers that opt in receive events with the same group ID in sequence. Other subscribers on the same bus still receive events asynchronously. - Subscriptions: a new resource that combines the event filter, target, retry policy, and dead-letter destination into one unit, replacing separate rules and targets.
- Content-based deduplication: EventBridge drops retried copies of the same payload within a five-minute window.
- JSONata transformations: subscribers can reshape events before delivery.
The pricing model changes too. Instead of per-event charges, publishers pay for ingress and subscribers pay for egress throughput. The default quota is 10,000 subscribers per bus. The feature launched in 15 regions across the US, Europe, and Asia Pacific. Ordering and dedup have long been the reasons teams put SQS FIFO or Kafka between EventBridge and consumers, so this could remove a layer from many architectures.
Read more — AWS News Blog
Google Previews GKE Agentic Migration: AI-Assisted EKS-to-GKE With Deterministic Guardrails
Google Cloud released GKE Agentic Migration in public preview. It is an open-source agent plugin, with an MCP server, that translates AWS EKS infrastructure-as-code and Kubernetes manifests into GKE landing zones. Google presents it as a replacement for ad-hoc LLM prompting. The output is reviewable pull requests, and the plugin never mutates a live cluster.
The design mixes LLM and deterministic steps. LLM workers write the more complex Terraform and YAML. Exact mappings, such as Workload Identity annotations and image registry rewrites, are handled by deterministic transforms. Everything is checked with terraform validate and Kubernetes manifest contract checks before a human approves it. The plugin maps AWS IRSA to Workload Identity, ALB ingress to the Gateway API, and Karpenter node claims to GKE Node Auto Provisioning or Custom Compute Classes.
Stateful data movement is kept out of scope on purpose. For databases and storage, the tool generates runbooks that point to Database Migration Service and Storage Transfer Service. A persistent migration state graph lets platform engineers approve the landing zone first. Application teams can then translate their own workloads in permission-isolated folders. The pattern of an LLM for authoring, deterministic validation, and a PR as the only way to apply changes is a useful template for any agent that touches infrastructure.
Read more — Google Cloud Blog
Links & Sources
You May Also Enjoy
