platformchangelog.dev
43 tools

Platform Changelog: what shipped in platform engineering, developer experience and DevOps

feedbefore 2026-10-26all releases, patches included
newest entries
2026-10-256 days ago
The New Stack: Is Kubernetes ready for AI agents?postThe New Stack argues that Kubernetes was built for software that does what its developers specify, and that AI agents strain that assumption. It says a CNCF survey found two-thirds of organizations running AI inference workloads do so on Kubernetes. The article calls for sandboxing, explicit permissioning, monitoring and warm pools for agents, and says the open-source project does not yet fully provide these.platform teamsPlatform teams running agent workloads should expect to need stronger sandboxing and permission controls than default Kubernetes offers today, and may need to consider commercial distributions with those features. The article describes an agent substrate using warm pools for fast agent startup but gives no implementation details or benchmarks.The New Stack ↗Amazon EKSwritten by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
2026-10-247 days ago
The New Stack: Building an agentic software factory? Start with the sensor.postThe article argues that agentic software factories are limited by their verification step, which it calls the sensor, rather than by coding agents. It defines three sensor properties (latency, fidelity, bandwidth) and suggests tracking time to signal, escape rate and concurrent verification capacity. The article is written by a team that sells a product for per-change realistic environments and presents that approach as the solution.platform teamsPlatform teams evaluating agent-driven pipelines can use the proposed metrics to find which verification property is weakest before raising agent autonomy. The article does not describe a specific tool change, and its recommended per-change environment approach comes from the vendor's own view.The New Stack ↗written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
The New Stack: Why Kubernetes autoscalers keep fighting each other (and why the solution must be autonomous)postThe article argues that Kubernetes autoscalers such as HPA, VPA, KEDA and Karpenter act on overlapping variables and can conflict, for example VPA shrinking requests that pushes HPA to scale out, or Argo CD reverting HPA replica changes. It describes four collision patterns and proposes a unified control layer that jointly decides replicas, resource requests and node capacity. The piece presents no new measured data beyond an illustrative example and a cited cluster CPU utilization figure of about 8%.platform teamsPlatform teams running HPA, VPA, KEDA or Karpenter together should check for overlapping writes to replicas and resource requests, and for GitOps auto-sync reverting autoscaler changes. The article offers no shipped tooling, so nothing changes in existing tools today.The New Stack ↗Argo CDFluxHelmKarpenterwritten by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
The New Stack: Before you expand your GPU pool, measure the outcomepostThe article argues that teams running AI workloads on Kubernetes should define outcomes, deadlines, and costs before expanding a GPU pool. It suggests staggering batch jobs with scheduling flexibility, load-testing sustainable serving capacity, and joining OpenCost allocation data with verified outcome records to compute cost per verified outcome.platform teamsPlatform teams can use OpenCost allocation data joined with serving telemetry and outcome records to decide between scheduling changes, batching, resizing, model changes, or new GPUs. The article notes OpenCost proposal 3915 for serving-engine telemetry is still a draft.The New Stack ↗written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
InfoQ: Google’s Data Agent Kit Hits GA, Splitting Agent Guardrails Between Skills and IAMpostGoogle made Data Agent Kit generally available, moving it out of its summer preview. The kit is open source and installs into coding agents such as Claude Code, Cursor and Codex, providing MCP tools for more than fifteen Google Data Cloud services plus 37 Google-authored skills under Apache 2.0. Google documents an IAM deny policy that blocks MCP tools where tool.isReadOnly is false, applied at organization, folder or project level.platform teamsPlatform teams running Google Cloud MCP servers can restrict agents to read-only tools with IAM deny policies instead of relying only on skill guidance. Blocked tools still appear in tools/list, the controls apply only to Google Cloud MCP servers, and the resource.service and tool.name attributes are available only through gcloud.InfoQ ↗written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
2026-10-238 days ago
Open Policy Agentv1.21.4patch
The New Stack: AI jobs are waiting while GPUs sit idle. Kubernetes has a new plan.postThe New Stack's Road to KubeCon column covers several Kubernetes topics: a HPE blog post on an open-source tool that gathers pod-restart diagnostics to Slack, an AWS post on scheduler and autoscaler friction for GPUs, and the llm-d v1.0 release. It also reports that in one benchmark by Skyhook's Radar, Claude Code given read-only MCP context diagnosed 50 injected faults faster than with shell and kubectl access, reaching a correct diagnosis within two minutes on 86% of faults versus 60%. Splunk's State of Observability 2026 survey of 2,050 professionals reports that over half cannot effectively measure AI quality or control AI costs.platform teamsPlatform teams evaluating AI troubleshooting agents may find that read-only MCP context improved speed in one author's benchmark, though the benchmark is a single small test by its author. The Splunk survey is vendor-commissioned, so its figures should be treated as indicative rather than verified.The New Stack ↗written by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
CNCF: Multi-tenant GPU Kubernetes on simulated hardware with Kamaji and HAMicase studyThe author built a single-machine lab where Kamaji runs two tenant Kubernetes control planes inside one management cluster and HAMi shares simulated GPUs between pods using a mock device plugin. The post reports that three fractional-GPU pods were scheduled onto one simulated A100 card, that HAMi refused pods exceeding the remaining cores or memory, and that eight tests passed. The source text provided is cut off before the tenant isolation test results are shown.platform teamsPlatform teams evaluating per-team Kubernetes isolation with GPU sharing can see a reproducible pattern for putting control planes in a shared cluster while scheduling fractional GPUs through HAMi. Results come from simulated hardware, so they do not show real GPU performance or driver behavior.CNCF Blog ↗Amazon EKSwritten by claude-haiku-5-5, prompt 7cbe154, checked by publish rules
daily email

One short email each morning: what shipped, what breaks, and a link to each change. Unsubscribe in one click. Privacy.

statuslast 30 days for lines
tools tracked43releases in registry964released in the last 30 days106entries published335breaking change lines18security lines84
change typestwo-column gutter, like git status
  • +addedNew functionality.
  • ~changedExisting behavior changed.
  • -deprecatedStill works, scheduled for removal.
  • -removedDecommissioned, gone in this version.
  • *fixedA defect corrected.
  • !securityA vulnerability closed or a dependency patched for one.
  • !breakingFlag in the first column; the change needs action before upgrading.
filter
categoryallidp & portalsai agents on platformci/cdobservabilitysupply chaincost & governancereports & dx dataenterprise app platforms
showmajor onlyreleasescase studiespostmortemsinclude patches