The Case for a Cloud Native Agent Harness
CNCF, Monday, September 28th, 2026
Craig McLuckie argues AI agent harnesses must be built as distributed applications, not desktop tools wrapped in containers.
Stacklok's Craig McLuckie argues that most coding-agent harnesses assume one user, one machine and one process that is simultaneously UI, agent loop, sandbox, credential store and session database, a design that breaks down when organizations need hundreds of governed sessions that survive node failures.
Applying Kubernetes' lesson that a monolith in a container is still a monolith, he proposes separating the agent loop from clients, execution environments, tool catalogs and supporting services.
The approach enables versioned rollouts, durable session state that survives worker loss, governed tool access and multiple client types, illustrated with the open source Mecatl harness.