In an article published on 28 October, Alibaba Cloud's native computing community lays out an 'Agent as a Service' design built on AgentScope. The argument is that an agent's identity, task state and accumulated memory should not belong to any single client, harness instance or execution environment, so that a personal agent can keep working after a user closes the laptop and resume the same task on return, and an enterprise agent can hand tool calls to cloud sandboxes or internal-network endpoints under its own permission rules. The proposed split: run the harness, the loop around the model, as a remote service; keep context, memory and recovery state in cloud state services; move execution ownership between local and remote through handoffs; and separate brain from hands so tool endpoints in different places can be swapped. The article lists four starting points - a code-first custom harness, a reused productised harness, a platform-hosted harness configured around a model, and a cloud product's built-in capabilities - and argues that agent definition, user session, inference instance and sandbox each follow their own lifecycle.
Our takeThe valuable part is the lifecycle split, not the diagram: an inference instance and a sandbox should expire on different clocks, which is exactly what most agent stacks conflate. But this is architecture without numbers - no resumption latency, no state-store failure modes. The testable questions are whether a task survives a node switch in the middle of a tool call, and whether tool execution is idempotent enough that a retried step does not act twice on the world.