The AI agent ecosystem is hitting a scaling wall, and Stacklok CEO Craig McLuckie has the blueprint to fix it. In a new post on the CNCF blog, McLuckie argues that the current definition of an "agent harness"β€”the bundle of context, filesystem, and permissions surrounding an AI agentβ€”is fundamentally flawed for production use. He contends that most existing harnesses are too local, designed to serve a single developer on a laptop rather than scaling to hundreds of concurrent sessions.

The Monolith in a Container

McLuckie draws a direct parallel between the early days of containerization and the current state of agentic AI. "Kubernetes taught this industry that a monolith in a container is still a monolith," he writes. The implication is stark: wrapping a tightly coupled agent loop in a container does not solve the architectural problems of state management and session persistence. To him, the solution is a distributed application that separates the agent loop from the underlying infrastructure and services, allowing sessions to move between clients and devices without breaking.

Kubernetes Scheduling Gaps Exposed

While agent architecture evolves, the underlying Kubernetes infrastructure is also showing cracks. A case study from Zhuoyu Technology, a Chinese autonomous driving firm, reveals that the default Kubernetes scheduler was capping GPU utilization. By adopting Koordinator, a CNCF sandbox project, Zhuoyu pushed GPU allocation above 95% and utilization above 55%. This highlights a critical reality for AI agents running on Kubernetes: without specialized schedulers, you are leaving massive performance on the table, regardless of how clever your agent harness is.

Agentic DevOps Goes Remote

The push for distributed, remote-capable agent interactions is gaining momentum beyond just the harness layer. Cycle, a DevOps control plane, recently launched a remote MCP (Model Context Protocol) server, allowing users to provision and manage multicloud workloads via natural language. Alexander Mattoni, Cycle’s head of engineering, described the capability as something that would have seemed like "science fiction" just a couple of years ago. This release follows a trend of abstracting DevOps operations into prompts, further validating the need for cloud-native agent interfaces that don't rely on local state.

Key Takeaways

  • Craig McLuckie argues that current agent harnesses are local monoliths that fail to scale across multiple sessions and devices.
  • Zhuoyu Technology achieved >95% GPU allocation by replacing the default Kubernetes scheduler with Koordinator.
  • Cycle released a remote MCP server, enabling natural language control of DevOps workflows across multicloud environments.
  • KubeCon + CloudNativeCon NA is approaching fast, running November 9-12 in Salt Lake City, with Argo CD 4.0 visioning beginning now.

The Bottom Line

If your AI agent can’t survive a laptop closing or a session timeout, you’re building a toy, not a product. The move to distributed, cloud-native harnesses isn't optional; it's the only way to escape the monolith trap.