Agent = Docker Container: A mental model that simplified everything in my agent infra project

The author, an AI researcher building agent infrastructure, found that modeling agents as Docker containers simplified key challenges around export, import, versioning, and distribution. After hitting blockers with workspace-based approaches in OpenClaw—particularly around data curation, secrets management, and runtime coupling—they pivoted to a runtime-agnostic abstraction layer using the mental model: agent = container, agent export = docker commit + push, agent import = docker pull + run. This reframing unlocked marketplace, versioning, rollback, and multi-runtime support using existing container tooling.