Google splits every AI agent session into an external harness that calls models and an internal sandbox microVM that executes shell commands, code, and browser actions