Amazon EKS adds seven-day control-plane rollback for Kubernetes upgrades

The single most consequential thing in this batch of AWS releases is boring and operational: you no longer have to rebuild an Amazon EKS cluster because an upgrade went sideways — at least for seven days.

Rollback window: right move, wrong clock

Amazon EKS now supports Kubernetes control-plane rollbacks within a seven-day window after an upgrade. Practically, that means if a Kubernetes upgrade introduces a regression in your cluster or breaks an operator, you can revert the control plane instead of spinning up a fresh cluster, migrating workloads, and stitching networking and RBAC back together. This fixes the most painful, time-consuming, and error-prone part of upgrading Kubernetes in AWS.

But the seven-day constraint matters. It will save you from flaky upgrades and third-party operator incompatibilities that show up immediately; it won't help with latent data-model migrations or operator changes that surface after your monitoring window. Treat it as a fast escape hatch, not a substitute for rigorous canary strategies and automated compatibility tests. If you want the operational playbook, shift fast-fail detection and rollback automation to the first 72 hours post-upgrade.

(If you want more context on the change, I covered the EKS rollback announcement recently.)

Lambda MicroVMs and 1 MB async payloads: bigger events, different primitives

AWS introduced Lambda MicroVMs as a lightweight VM-style serverless primitive and increased the maximum asynchronous invocation payload to 1 MB (previously 256 KB), with coordinated support promised across async event services. Two linked implications:

MicroVMs acknowledge that container runtimes aren't the only right abstraction for serverless. A small VM primitive improves startup isolation and can support resumable execution patterns and stronger tenant separation. Expect questions about observability and cold-start characteristics, but this is overdue.
The 1 MB async payload change matters: you can now send richer context with events without immediately resorting to external payload stores. That simplifies many event-driven designs but also enlarges the implicit trust and data footprint in async workflows. Instrument and audit your event consumers accordingly.

My take: MicroVMs are the right call. But platform teams who treat them as "just another runtime" will pay for it in tracing, security posture, and cost attribution unless they bake new conventions into observability and CI/CD.

AgentCore GA: agents, web search, and the expanding trust boundary

Amazon Bedrock AgentCore reached general availability with tooling for retrieval-augmented agents, tool invocation, and web search integration. This turns agents from experiments into a managed runtime: a framework for knowledge-grounded agents that can call tools, run searches, and integrate with external systems.

The functional upside is obvious — simpler builder velocity for retrieval-augmented agents and tool-using assistants. The security and governance angle is the thing platform teams should care about: agents with tool invocation and web search are a new attack surface. Audit logs, least-privilege tool connectors, and limits on emergent capability need to be part of any AgentCore rollout. This is not a feature you bolt on to existing IAM models and expect to be safe.

Graviton instances and the miscellaneous stack

AWS also announced new Graviton-based instance families for C- and M-class workloads and positioned them as a performance upgrade over prior generations. For compute-heavy workloads and inference hosts that are already Arm-native, this is an attractive migration path.

AWS added ACME support to ACM and expanded CloudFormation's Express Mode for faster, AI-driven template generation — small ergonomics wins that nudge teams toward more automation.

What to change tomorrow

Treat EKS rollbacks as an emergency escape hatch and automate detection/rollback in the first 72 hours. Don't let it become your testing strategy.
If you plan to use Lambda MicroVMs, define tracing, cost, and security conventions up front. MicroVM != opaque black box.
AgentCore GA is an ops and governance problem, not just a developer convenience. Lock down tool connectors and add telemetry for every agent action.

This release slate is quietly architectural. AWS reduced operational drag (EKS), changed runtime primitives (Lambda MicroVMs), enlarged event payloads (1 MB async), and pushed managed agent tooling into production (AgentCore). Each change improves velocity — but also shifts responsibilities: platform teams now need better early-detection runbooks, stronger agent governance, and observability that understands VM-level serverless. Ignore that, and you'll get faster deployments and slower incident responses.

Amazon EKS adds seven-day control-plane rollback for Kubernetes upgrades

Sources

AWS Lambda MicroVMs and 1 MB Async Payloads: What Platform Teams Should Know

AWS Lambda MicroVMs: VM-level isolation and resumable state for serverless agents

Amazon EKS: Control-Plane Rollbacks Supported Within a Seven-Day Window