Choose Your AI Agent Host Based on Data You Can Actually Export

A developer argues that the choice of where to run an AI agent should be driven by whether you can fully export the run's artifacts — including the prompt file, tool-call log, workspace hashes, and a record of refused commands. Without these four elements forming a portable 'exit bundle,' the author treats any host as a temporary sketchpad rather than a reliable workspace. Three hosting options are compared: free shared servers, self-hosted machines, and paid isolated environments, each suited to different levels of ownership and auditability needs. The core concern is not cost but control — specifically, who retains access to the workspace after a session ends and whether a teammate could reconstruct the agent's reasoning without relying on chat history. The author frames this as a practical pre-demo checklist borrowed from code-review discipline, warning that a free host used for a quick spike can quietly become the permanent and unauditable home for important work.