AX and OpenShell manage execution environments, but neither maps discussions to agent prompt turns. AX has Task, Workspace, and Model resources; OpenShell manages sandbox lifecycle and policy. Their APIs still need an adapter for follow-ups, result routing, and runtime cancellation.

Runtime controllers cover another part of the contract

Google AX declares a Task with a sandboxed command, Workspaces with repositories/MCP/skills, and Model bindings. It exposes task create/watch/suspend/resume/delete. AX explicitly does not model an agent's planning, delegation, and fan-out. Its README warns of breaking changes before a stable release. Core concepts, repository README.

The current AX runner retains workspace files across suspend/resume but starts a fresh container and process tree. It also does not yet report its agent command's exit status to the control plane. Completion reporting and prompt-turn delivery need an agent protocol or application callback. Budgets and approval policies appear in AX's roadmap, not a verified current feature. Runner contract, roadmap.

NVIDIA OpenShell's SDK can create, observe, exec into, stop/start, and delete a policy-governed sandbox. It does not define the human task or agent conversation. Python SDK.

The proposed integration stores runId -> runtimeId, uses an agent-facing protocol such as ACP or a wrapper to deliver turns, and verifies that a requested stop actually ended the concrete runtime. This connects to the prior fleet assessment. It is not a tested AX or OpenShell deployment.