Cloud-Native Agent Harnesses Must Break Free From Single-Developer Designs

The infrastructure surrounding AI agents—their context, permissions, services, and supporting systems—has become increasingly complex. Yet most current frameworks remain tethered to a single-developer laptop model, struggling to scale across multiple concurrent sessions or move seamlessly between devices.

Stacklok's Craig McLuckie contends that this local-first approach is fundamentally limiting. He points to a critical architectural lesson from the Kubernetes ecosystem: containerizing a monolith doesn't make it less monolithic. That insight, he argues, applies directly to AI agent harnesses.

The Case for Distributed Agent Architecture#

Rather than bundling everything into a single application, cloud-native agent harnesses should separate the agent's core loop from the infrastructure and services supporting it. This separation allows sessions to persist across devices, scales to hundreds of simultaneous users, and reduces failures that leave agents stranded when individual components break. McLuckie elaborated on this vision on the CNCF blog, framing the shift as a necessary evolution as agents move from development environments into production systems.

Other Infrastructure Advances#

The broader ecosystem continues evolving to meet these demands. Kubernetes scheduler improvements are enabling dramatic gains in GPU utilization—Koordinator, a CNCF sandbox project, helped autonomous driving company Zhuoyu Technology push GPU allocation above 95% and utilization above 55%, far exceeding what the default Kubernetes scheduler achieves. Meanwhile, platforms like Cycle are adding natural-language interfaces for multicloud orchestration, bringing agentic capabilities directly into DevOps workflows.

These advances signal a shift: the infrastructure layer itself is becoming smarter, more distributed, and increasingly agent-aware.