The New Stack’s sponsored Road to KubeCon roundup examines how AI agent harnesses may need to evolve from laptop-oriented tools into cloud-native distributed applications. Stacklok founder and CEO Craig McLuckie argues that a harness, which coordinates an agent’s context, filesystem, subagents, permissions, and related services, should separate the agent loop from the infrastructure around it. He compares the current approach with Kubernetes’ lesson that putting a monolith in a container does not make it distributed; current harnesses can struggle with hundreds of sessions, failures, and movement between clients or devices. The article also reports a Zhuoyu Technology case study using Koordinator, a CNCF sandbox project, to schedule autonomous-driving workloads on Kubernetes: GPU allocation exceeded 95% and overall GPU utilization exceeded 55%, after limitations in the default scheduler contributed to failed launches, stranded GPUs, and distributed-job problems. Other items include a 30-day CNCF and OpenSSF Security Slam running October 5 through November 6, and Atlassian’s AutoHOT detection platform, which combines OpenTelemetry, Kafka, and Flink on Kubernetes to reduce event-to-metric latency from more than 40 seconds to under 10 seconds across telemetry from more than 10 cloud products and millions of tenants; recall was 64% in August and tuning continues. The roundup previews ArgoCon North America and the early visioning process for Argo CD 4.0. It also describes Cycle’s remote MCP server, which lets users provision, orchestrate, and manage workloads across multicloud and hybrid environments through compatible AI assistants and coding tools. The piece is part of an HPE-presented series leading up to KubeCon + CloudNativeCon North America in Salt Lake City on November 9-12, 2026.
