AI patches cheap, meetings cost: K8s breaks at 100k nodes

When generative AI writes a trivial change in minutes, the bottleneck shifts to the discussion about whether to ship it. Using the AI‑generated diff as a concrete probe lets teams decide in half an hour instead of days, keeping scope‑creep debates from draining resources.

GitLab’s Duo Agent Platform adds Custom and Foundational AI flows that trigger from native events and execute multi‑step tasks, like fixing pipelines or reviewing merge requests, without manual handoffs. Teams can embed human‑in‑the‑loop approvals, enforce governance, and reuse trusted sequences across projects, turning runbooks into production‑ready automation.

HAMi, the open-source GPU virtualization layer for Kubernetes, entered CNCF incubation after attracting over 2.6 k contributors and 550 organizations. It lets teams slice GPUs by memory or cores, enforce isolation, and schedule workloads without code changes, paving the way for broader, vendor-neutral AI infrastructure on K8s.

Running Amazon EKS’s Network Policy and VPC Resource controllers across clusters up to a hundred‑thousand nodes exposed the hidden cost of intent‑to‑enforcement gaps. Cache lag, memory pressure and churn‑driven reconciliation errors turned into reliability failures, prompting concrete patterns for scaling controllers safely.

Uber merged OpenSearch’s shard‑allocation awareness with its own isolation‑group system to keep search and ingest alive even if an entire zone disappears. By forcing shard placement and spreading nodes across three isolation groups, a cluster with three copies and five manager nodes survives a zone loss plus a node failure without downtime.

DoorDash just launched “dd‑cli”, a command‑line tool that lets autonomous AI agents browse restaurants, add items, and pay without human approval. By exposing its ordering backend, DoorDash is betting that platform‑level APIs will become the next battleground for AI‑driven commerce, potentially reshaping its revenue model.

At GTC 2026 NVIDIA unveiled OpenShell, an open‑source sandbox that enforces kernel‑level policies on AI agents, controlling file access, process execution, and network traffic. By moving security to the environment, it prevents compromised agents from leaking credentials or bypassing restrictions, complementing Tigera Lynx for cross‑sandbox governance.

Datadog built a suite of MCP Server tools that let agents handle rule authoring, raw signal inspection, and bulk triage inside Cloud SIEM. By mining usage data, using progressive disclosure to protect a shared context window, and a custom eval framework to test non‑deterministic behavior, the toolset stays reliable as teams grow.